Systematic discovery of directional regulatory motifs associated with human insulator sites

  1. Naoki Osato
  2. Michiaki Hamada

Reviewed by

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Insulator proteins function as barriers to enhancer–promoter interactions (EPIs), thereby regulating gene expression. The primary insulator protein in vertebrates is CTCF, a DNA-binding protein (DBP); however, the roles of other DBPs in EPI insulation are not fully understood. To address this, we developed a systematic and comprehensive deep learning–based approach to identify DNA motifs of DBPs associated with insulator function. Applying this method to human fibroblast cells, we identified 97 directional motifs and a smaller number of non-directional motifs. These motifs were mapped to 23 DBPs previously linked to insulator activity, CTCF, and/or other forms of chromosomal transcriptional regulation. We found that the estimated orientation bias of CTCF was consistently proportional to the orientation bias observed in chromatin interaction data. Furthermore, these motifs showed significant enrichment at insulator sites that separate repressive and active chromatin regions, at chromatin interaction–defined boundaries, and at splice sites, compared to motifs of other DBPs. For instance, we observed that the key regulator MyoD-binding site is located at an insulator site near a gene involved in skeletal muscle differentiation and function. Importantly, our findings support the previously proposed insulator-pairing model, which suggests that insulator–insulator interactions are orientation-dependent, and highlight the involvement of multiple DNA-binding proteins beyond CTCF. Together, these results provide new insights into transcriptional regulatory mechanisms mediated by insulator-associated DBPs.

Article activity feed

  1. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Reply to the Reviewers

    I would like to thank the reviewers for their comments and interest in the manuscript and the study.

    Reviewer #1

    1. I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning.

    The directional positioning of CTCF-binding sites at chromatin interaction sites was analyzed by CRISPR experiment (Guo Y et al. Cell 2015). We found that the machine learning and statistical analysis showed the same directional bias of CTCF-binding motif sequence and RAD21-binding motif sequence at chromatin interaction sites as the experimental analysis of Guo Y et al. (lines 229-253, Figure 3b, c, d and Table 1). Since CTCF is involved in different biological functions (Braccioli L et al. Essays Biochem. 2019 ResearchGate webpage), the directional bias of binding sites may be reduced in all binding sites including those at chromatin interaction sites (lines 68-73). In our study, we investigated the DNA-binding sites of proteins using the ChIP-seq data of DNA-binding proteins and DNase-seq data. We also confirmed that the DNA-binding sites of SMC3 and RAD21, which tend to be found in chromatin loops with CTCF, also showed the same directional bias as CTCF by the computational analysis.

    __2. Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. __

    Following the reviewer's advice, I performed the same analysis with the DNA-binding sites that do no overlap with the DNA-binding sites of CTCF and cohesin (RAD21 and SMC3) (Fig. 6 and Supplementary Fig. 4). The result showed the same tendency in the distribution of DNA-binding sites. The height of a peak on the graph became lower for some DNA-binding proteins after removing the DNA-binding sites that overlapped with those of CTCF and cohesin. I have added the following sentence on lines 435 and 829: For the insulator-associated DBPs other than CTCF, RAD21, and SMC3, the DNA-binding sites that do not overlap with those of CTCF, RND21, and SMC3 were used to examine their distribution around interaction sites.

    3. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As suggested by the reviewer, I have added the insulator scores and boundary sites from the 4D nucleome data portal as tracks in the UCSC genome browser. The insulator scores seem to correspond to some extent to the H3K27me3 histone marks from ChIP-seq (Fig. 4a and Supplementary Fig. 3). We found that the DNA-binding sites of the insulator-associated DBPs were statistically overrepresented in the 5 kb boundary sites more than other DBPs (Fig. 4d). The direction of DNA-binding sites on the genome can be shown with different colors (e.g. red and green), but the directionality of insulator-associated DNA-binding sites is their overall tendency, and it may be difficult to notice the directionality from each binding site because the directionality may be weaker than that of CTCF, RAD21, and SMC3 as shown in Table 1 and Supplementary Table 2. We also observed the directional biases of CTCF, RAD21, and SMC3 by using Micro-C chromatin interaction data as we estimated, but the directionality was more apparent to distinguish the differences between the four directions of FR, RF, FF, and RR using CTCF-mediated ChIA-pet chromatin interaction data (lines 287 and 288).

     I found that the CTCF binding sites examined by a wet experiment in the previous study may not always overlap with the boundary sites of chromatin interactions from Micro-C assay (Guo Y et al. *Cell* 2015). The chromatin interaction data do not include all interactions due to the high sequencing cost of the assay, and include less long-range interactions due to distance bias. The number of the boundary sites may be smaller than that of CTCF binding sites acting as insulators and/or some of the CTCF binding sites may not be locate in the boundary sites. It may be difficult for the boundary location algorithm to identify a short boundary location. Due to the limitations of the chromatin interaction data, I planned to search for insulator-associated DNA-binding proteins without using chromatin interaction data in this study.

 I discussed other causes in lines 614-622: Another reason for the difference may be that boundary sites are more closely associated with topologically associated domains (TADs) of chromosome than are insulator sites. Boundary sites are regions identified based on the separation of numerous chromatin interactions. On the other hand, we found that the multiple DNA-binding sites of insulator-associated DNA-binding proteins were located close to each other at insulator sites and were associated with distinct nested and focal chromatin interactions, as reported by Micro-C assay. These interactions may be transient and relatively weak, such as tissue/cell type, conditional or lineage-specific interactions.

 Furthermore, I have added the statistical summary of the analysis in lines 372-395 as follows: Overall, among 20,837 DNA-binding sites of the 97 insulator-associated proteins found at insulator sites identified by H3K27me3 histone modification marks (type 1 insulator sites), 1,315 (6%) overlapped with 264 of 17,126 5kb long boundary sites, and 6,137 (29%) overlapped with 784 of 17,126 25kb long boundary sites in HFF cells. Among 5,205 DNA-binding sites of the 97 insulator-associated DNA-binding proteins found at insulator sites identified by H3K27me3 histone modification marks and transcribed regions (type 2 insulator sites), 383 (7%) overlapped with 74 of 17,126 5-kb long boundary sites, 1,901 (37%) overlapped with 306 of 17,126 25-kb long boundary sites. Although CTCF-binding sites separate active and repressive domains, the limited number of DNA-binding sites of insulator-associated proteins found at type 1 and 2 insulator sites overlapped boundary sites identified by chromatin interaction data. Furthermore, by analyzing the regulatory regions of genes, the DNA-binding sites of the 97 insulator-associated DNA-binding proteins were found (1) at the type 1 insulator sites (based on H3K27me3 marks) in the regulatory regions of 3,170 genes, (2) at the type 2 insulator sites (based on H3K27me3 marks and gene expression levels) in the regulatory regions of 1,044 genes, and (3) at insulator sites as boundary sites identified by chromatin interaction data in the regulatory regions of 6,275 genes. The boundary sites showed the highest number of overlaps with the DNA-binding sites. Comparing the insulator sites identified by (1) and (3), 1,212 (38%) genes have both types of insulator sites. Comparing the insulator sites between (2) and (3), 389 (37%) genes have both types of insulator sites. From the comparison of insulator and boundary sites, we found that (1) or (2) types of insulator sites overlapped or were close to boundary sites identified by chromatin interaction data.

    4. The suggested alternative transcripts function, also highlighted in the manuscripts abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.

    According to the reviewer's comment, I performed the genome-wide analysis of alternative transcripts where the DNA-binding sites of insulator-associated proteins are located near splicing sites. The DNA-binding sites of insulator-associated DNA-binding proteins were found within 200 bp centered on splice sites more significantly than the other DNA-binding proteins (Fig. 4e and Table 2). I have added the following sentences on lines 405 - 412: We performed the statistical test to estimate the enrichment of insulator-associated DNA-binding sites compared to the other DNA-binding proteins, and found that the insulator-associated DNA-binding sites were significantly more abundant at splice sites than the DNA-binding sites of the other proteins (Fig 4e and Table 2; Mann‒Whitney U test, p value 5. Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.

    I believe that the Figure 1 would help researchers in other fields who are not familiar with biological phenomena and functions to understand the study. More explanation has been included in the Figures and legends of Figs. 4 and 5 to help readers outside the immediate research field understand the figures.

    6. Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Reviewer #2's comments would be related to this comment. I have introduced a more detailed explanation of the method in the Results section, as shown in the responses to Reviewer #2's comments.

    Reviewer #2

    1. Introduction, line 95: CTCF appears two times, it seems redundant.

    On lines 91-93, I deleted the latter CTCF from the sentence "We examine the directional bias of DNA-binding sites of CTCF and insulator-associated DBPs, including those of known DBPs such as RAD21 and SMC3".

    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?

    Although CTCF is known to be the main insulator protein in vertebrates, we found that 97 DNA-binding proteins including CTCF and cohesin are associated with insulator sites by modifying and developing a machine learning method to search for insulator-associated DNA-binding proteins. Most of the insulator-associated DNA-binding proteins showed the directional bias of DNA-binding motifs, suggesting that the directional bias is associated with the insulator.

     I have added the sentence in lines 96-99 as follows: Furthermore, statistical testing the contribution scores between the directional and non-directional DNA-binding sites of insulator-associated DBPs revealed that the directional sites contributed more significantly to the prediction of gene expression levels than the non-directional sites. I have revised the statement in lines 101-110 as follows: To validate these findings, we demonstrate that the DNA-binding sites of the identified insulator-associated DBPs are located within potential insulator sites, and some of the DNA-binding sites in the insulator site are found without the nearby DNA-binding sites of CTCF and cohesin. Homologous and heterologous insulator-insulator pairing interactions are orientation-dependent, as suggested by the insulator-pairing model based on experimental analysis in flies. Our method and analyses contribute to the identification of insulator- and chromatin-associated DNA-binding sites that influence EPIs and reveal novel functional roles and molecular mechanisms of DBPs associated with transcriptional condensation, phase separation and transcriptional regulation.

    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS.

    On lines 121-124, to explain the procedure for the SNP of an eQTL, I have added the sentence in the Methods: "If a DNA-binding site was located within a 100-bp region around a single-nucleotide polymorphism (SNP) of an eQTL, we assumed that the DNA-binding proteins regulated the expression of the transcript corresponding to the eQTL".

    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details.

    On line 119, I have included the explanation of the eQTL dataset of GTEx v8 as follows: " The eQTL data were derived from the GTEx v8 dataset, after quality control, consisting of 838 donors and 17,382 samples from 52 tissues and two cell lines". On lines 681 and 865, I have added the filename of the eQTL data "(GTEx_Analysis_v8_eQTL.tar)".

    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.

    The reviewer would mention Figure 2, not Figure 1. If so, the matrices in panels a and b in Figure 2 are equivalent. I have shown it in the figure: The same figure in panel a is rotated 90 degrees to the right. The green boxes in the matrix show the regions with the ChIP-seq peak of a DNA-binding protein overlapping with a SNP of an eQTL. I used eQTL data to associate a gene with a ChIP-seq peak that was more than 2 kb upstream and 1 kb downstream of a transcriptional start site of a gene. For each gene, the matrix was produced and the gene expression levels in cells were learned and predicted using the deep learning method. I have added the following sentences to explain the method in lines 133 - 139: Through the training, the tool learned to select the binding sites of DNA-binding proteins from ChIP-seq assays that were suitable for predicting gene expression levels in the cell types. The binding sites of a DNA-binding protein tend to be observed in common across multiple cell and tissue types. Therefore, ChIP-seq data and eQTL data in different cell and tissue types were used as input data for learning, and then the tool selected the data suitable for predicting gene expression levels in the cell types, even if the data were not obtained from the same cell types.

    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?

    As suggested by the reviewer, to help readers understand the observation, I have added Supplementary Fig. S4c to show the distribution of DNA-binding sites of "CTCF, RAD21, and SMC3" and "BACH2, FOS, ATF3, NFE2, and MAFK" around chromatin interaction sites. I have modified the following sentence to indicate the figure on line 501: Although a DNA-binding-site distribution pattern around chromatin interaction sites similar to those of CTCF, RAD21, and SMC3 was observed for DBPs such as BACH2, FOS, ATF3, NFE2, and MAFK, less than 1% of the DNA-binding sites of the latter set of DBPs colocalized with CTCF, RAD21, or SMC3 in a single bin (Fig. S4c).

     In Aljahani A et al. *Nature Communications* 2022, we find that depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Together, our data show that loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression. Goel VY et al. *Nature Genetics* 2023 mentioned in the abstract: Microcompartments frequently connect enhancers and promoters and though loss of loop extrusion and inhibition of transcription disrupts some microcompartments, most are largely unaffected. These results suggested that chromatin loops can be driven by other DBPs independent of the known CTCF/Cohesin.

I added the following sentence on lines 569-577: The depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. Furthermore, the loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression.

 FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates (Ji D et al. *Molecular Cell* 2024). CTCF have also found to form transcriptional condensate and phase separation (Lee R et al. *Nucleic acids research* 2022). FOS was found to be an insulator-associated DNA-binding protein in this study and is potentially involved in chromatin remodeling, transcription condensation, and phase separation with the other factors such as BACH2, ATF3, NFE2 and MAFK. I have added the following sentence on line 556: FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates.

    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?

    Goel VY et al. Nature Genetics 2023 identified highly nested and focal interactions through region capture Micro-C, which resemble fine-scale compartmental interactions and are termed microcompartments. In the section titled "Most microcompartments are robust to loss of loop extrusion," the researchers noted that a small proportion of interactions between CTCF and cohesin-bound sites exhibited significant reductions in strength when cohesin was depleted. In contrast, the majority of microcompartmental interactions remained largely unchanged under cohesin depletion. Our findings indicate that most P-P and E-P interactions, aside from a few CTCF and cohesin-bound enhancers and promoters, are likely facilitated by a compartmentalization mechanism that differs from loop extrusion. We suggest that nested, multiway, and focal microcompartments correspond to small, discrete A-compartments that arise through a compartmentalization process, potentially influenced by factors upstream of RNA Pol II initiation, such as transcription factors, co-factors, or active chromatin states. It follows that if active chromatin regions at microcompartment anchors exhibit selective "stickiness" with one another, they will tend to co-segregate, leading to the development of nested, focal interactions. This microphase separation, driven by preferential interactions among active loci within a block copolymer, may account for the striking interaction patterns we observe.

     The authors of the paper proposed several mechanisms potentially involved in microcompartments. These mechanisms may be involved in looping with insulator function. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently (Hsieh TS et al. *Nature Genetics* 2022). Among the identified insulator-associated DNA-binding proteins, Maz and MyoD1 form loops without CTCF (Xiao T et al. *Proc Natl Acad Sci USA* 2021 ; Ortabozkoyun H et al. *Nature genetics* 2022 ; Wang R et al. *Nature communications* 2022). I have added the following sentences on lines 571-575: Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. I have included the following explanation on lines 582-584: Maz and MyoD1 among the identified insulator-associated DNA-binding proteins form loops without CTCF.

 As for the directionality of CTCF, if chromatin loop anchors have some structural conformation, as shown in the paper entitled "The structural basis for cohesin-CTCF-anchored loops" (Li Y et al. *Nature* 2020), directional DNA binding would occur similarly to CTCF binding sites. Moreover, cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops (Davidson IF et al. *Nature Reviews Molecular Cell Biology* 2021). Regarding loop extrusion, the 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions (Guerin TM et al. *EMBO Journal* 2024). I have added the following sentences on lines 543-547: Cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops. I have included the following sentences on lines 577-582: The 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions.

 Another model for the regulation of gene expression by insulators is the boundary-pairing (insulator-pairing) model (Bing X et al. *Elife* 2024) (Ke W et al. *Elife* 2024) (Fujioka M et al. *PLoS Genetics* 2016). Molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies. Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent. I have summarized the model on lines 559-567: Other types of chromatin regulation are also expected to be related to the structural interactions of molecules. As the boundary-pairing (insulator-pairing) model, molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies (Fig. 7). Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent.

    8. Do the authors think that the identified DBPs could work in that way as well?

    The boundary-pairing (insulator-pairing) model would be applied to the insulator-associated DNA-binding proteins other than CTCF and cohesin that are involved in the loop extrusion mechanism (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016).

     Liquid-liquid phase separation was shown to occur through CTCF-mediated chromatin loops and to act as an insulator (Lee, R et al. *Nucleic Acids Research* 2022). Among the identified insulator-associated DNA-binding proteins, CEBPA has been found to form hubs that colocalize with transcriptional co-activators in a native cell context, which is associated with transcriptional condensate and phase separation (Christou-Kent M et al. *Cell Reports* 2023). The proposed microcompartment mechanisms are also associated with phase separation. Thus, the same or similar mechanisms are potentially associated with the insulator function of the identified DNA-binding proteins. I have included the following information on line 554: CEBPA in the identified insulator-associated DNA-binding proteins was also reported to be involved in transcriptional condensates and phase separation.

    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?

    Snead WT et al. Molecular Cell 2019 mentioned that protein post-transcriptional modifications (PTMs) facilitate the control of molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin (Tang X et al. Nature Communications 2024). I found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Supplementary Fig. 2d). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation by PTMs. I have added the following explanation on lines 584-590: Furthermore, protein post-transcriptional modifications (PTMs) facilitate control over the molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin. We found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Fig. 4f and Supplementary Fig. 3c). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation through PTMs.

    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Structural molecular model of cohesin-CTCF-anchored loops has been published by Li Y et al. Nature 2020. The structural conformation of CTCF and cohesin in the loops would be the cause of the directional bias of CTCF binding sites, which I mentioned in lines 539 - 543 as follows: These results suggest that the directional bias of DNA-binding sites of insulator-associated DBPs may be involved in insulator function and chromatin regulation through structural interactions among DBPs, other proteins, DNAs, and RNAs. For example, the N-terminal amino acids of CTCF have been shown to interact with RAD21 in chromatin loops.

     To investigate the principles underlying the architectural functions of insulator-insulator pairing interactions, two insulators, Homie and Nhomie, flanking the *Drosophila even skipped *locus were analyzed. Pairing interactions between the transgene Homie and the eve locus are directional. The head-to-head pairing between the transgene and endogenous Homie matches the pattern of activation (Fujioka M et al. *PLoS Genetics* 2016).

    Reviewer #3

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.

    When a protein complex binds to DNA, one protein of the complex binds to the DNA directory, and the other proteins may not bind to DNA. However, the DNA motif sequence bound by the protein may be registered as the DNA-binding motif of all the proteins in the complex. The molecular structure of the complex of CTCF and Cohesin showed that both CTCF and Cohesin bind to DNA (Li Y et al. Nature 2020). I think there is a possibility that if the molecular structure of a protein complex becomes available, the previous recognition of the DNA-binding ability of a protein may be changed. Therefore, I searched the Pfam database for 99 insulator-associated DNA-binding proteins identified in this study. I found that 97 are registered as DNA-binding proteins and/or have a known DNA-binding domain, and EP300 and SIN3A do not directory bind to DNA, which was also checked by Google search. I have added the following explanation in line 257 to indicate direct and indirect DNA-binding proteins: Among 99 insulator-associated DBPs, EP300 and SIN3A do not directory interact with DNA, and thus 97 insulator-associated DBPs directory bind to DNA. I have updated the sentence in line 20 of the Abstract as follows: We discovered 97 directional and minor nondirectional motifs in human fibroblast cells that corresponded to 23 DBPs related to insulator function, CTCF, and/or other types of chromosomal transcriptional regulation reported in previous studies.

    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.

    As the reviewer mentioned, I recognize enhancers are relatively small regions. In the paper, I intended to examine further upstream and downstream of promoter regions where enhancers are found. Therefore, I have modified the sentence in lines 929 - 931 of the Fig. 2 legend as follows: Enhancer-gene regulatory interaction regions consist of 200 bins of 10 kbp between -1 Mbp and 1 Mbp region from TSS, not including promoter.

    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.

    Following the reviewer's advice, I have added the ChIP-seq data of H3K9me3 as a truck of the UCSC Genome Browser. The distribution of H3K9me3 signal was different from that of H3K27me3 in some regions. I also found the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions and took some screenshots of the UCSC Genome Browser of the regions around the sites in Supplementary Fig. 3b. I have modified the following sentence on lines 974 - 976 in the legend of Fig. 4: a Distribution of histone modification marks H3K27me3 (green color) and H3K9me3 (turquoise color) and transcript levels (pink color) in upstream and downstream regions of a potential insulator site (light orange color). I have also added the following result on lines 356 - 360: The same analysis was performed using H3K9me3 marks, instead of H3K27me3 (Fig. S3b). We found that the distribution of H3K9me3 signal was different from that of H3K27me3 in some regions, and discovered the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions (Fig. S3b).

    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    The resolution of the Micro-C assay is considered to be 100 bp and above, as the human nucleome core particle contains 145 bp (and 193 bp with linker) of DNA. However, internucleosomal DNA is cleaved by endonuclease into fragments of multiples of 10 nucleotides (Pospelov VA et al. *Nucleic Acids Research *1979). Highly nested focal interactions were observed (Goel VY et al. *Nature Genetics *2023). Base pair resolution was reported using Micro Capture-C (Hua P et al. Nature 2021). Sub-kilobase (20 bp resolution) chromatin topology was reported using an MNase-based chromosome conformation capture (3C) approach (Aljahani A et al. *Nature Communications *2022). On the other hand, Hi-C data was analyzed at 1 kb resolution. (Gu H et al. bioRxiv 2021). If the resolution of Micro-C interactions is at best at 1 kb, the binding sites of a DNA-binding protein will not show a peak around the center of the genomic locations of interaction edges. Each panel shows the number of binding sites of a specific DNA-binding protein at a specific distance from the midpoint of all chromatin interaction edges. I have modified and added the following sentences in lines 593-597: High-resolution chromatin interaction data from a Micro-C assay indicated that most of the predicted insulator-associated DBPs showed DNA-binding-site distribution peaks around chromatin interaction sites, suggesting that these DBPs are involved in chromatin interactions and that the chromatin interaction data has a high degree of resolution. Base pair resolution was reported using Micro Capture-C.

    Minor Comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., ____https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2____or ____https://pubmed.ncbi.nlm.nih.gov/37486787____/). The authors should discuss how that would impact their results.

    The directional bias of CTCF binding sites was identified by ChIA-pet interactions of CTCF binding sites. The analysis of the contribution scores of DNA-binding sites of proteins considering the binding sites of CTCF as an insulator showed the same tendency of directional bias of CTCF binding sites. In the analysis, to remove the false-positive prediction of DNA-binding sites, I used the binding sites that overlapped with a ChIP-seq peak of the DNA-binding protein. This result suggests that the DNA-binding sites of CTCF obtained by the current analysis have sufficient quality. Therefore, if the accuracy of prediction of DNA-binding sites is improved, although the number of DNA-binding sites may be different, the overall tendency of the directionality of DNA-binding sites will not change and the results of this study will not change significantly.

     As for the first reference in the reviewer's comment, chromatin interaction data from Micro-C assay does not include all chromatin interactions in a cell or tissue, because it is expensive to cover all interactions. Therefore, it would be difficult to predict all chromatin interactions based on machine learning. As for the second reference in the reviewer's comment, pioneer factors such as FOXA are known to bind to closed chromatin regions, but transcription factors and DNA-binding proteins involved in chromatin interactions and insulators generally bind to open chromatin regions. The search for the DNA-binding motifs is not required in closed chromatin regions.

    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    In the DeepLIFT paper, the authors explain that DeepLIFT is a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input (Shrikumar A et al. ICML 2017). DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

     Truly explainable AI would be able to find cause and reason, and to make choices and decisions like humans. DeepLIFT does not perform causal inferences. I did not use the term "Explainable AI" in our manuscript, but I briefly explained it in Discussion. I have added the following explanation in lines 623-628: AI (Artificial Intelligence) is considered as a black box, since the reason and cause of prediction are difficult to know. To solve this issue, tools and methods have been developed to know the reason and cause. These technologies are called Explainable AI. DeepLIFT is considered to be a tool for Explainable AI. However, DeepLIFT does not answer the reason and cause for a prediction. It calculates scores representing the contribution of the input data to the prediction.

 Furthermore, to improve the readability of the manuscript, I have included the following explanation in lines 159-165: we computed DeepLIFT scores of the input data (i.e., each binding site of the ChIP-seq data of DNA-binding proteins) in the deep leaning analysis on gene expression levels. DeepLIFT compares the importance of each input for predicting gene expression levels to its 'reference or background level' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.
  2. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary:

    Osato and Hamada propose a systematic approach to identify DNA binding proteins that display directional binding. They used a modified Deep Learning method (DEcode) to investigate binding profiles of 1356 DBP from GTRD database at promoters (30 of 100bp bins around TSS) and enhancers (200 bins of 10Kb around eSNPs) and use this to predict expression of 25,071 genes in Fibroblasts, Monocytes, HMEC and NPC. This method achieves a good prediction power (Spearman correlation between predicted and actual expression of 0.74). They then use PIQ, and overlap predicted binding sites with actual ChIP-seq data to investigate the motifs of TFs that are controlling gene expression. They find 99 insulator proteins showing either a specific directional bias or minor non-directional bias, corresponding to 23 DBP previously reported to have insulator function. Of the 23 proteins they identify as regulating enhancer promoter interactions, 13 are associated with CTCF. They also show that there are significantly more insulator proteins binding sites at borders of polycomb domains, transcriptionally active or boundary regions based on chromatin interactions than other proteins.

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.
    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.
    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.
    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    Minor comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.
    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    Referee Cross-Commenting

    I would like to mention that I agree with the comments of reviewers 1 and 2.

    Significance

    General assessment:

    This is the first study to my knowledge that attempts to use Deep Learning to identify insulators and directional biases in binding. One of the limitations is that no additional methods were used to show that these DBP have directional binding bias. It is not necessarily to employ additional methods, but it would definitely strengthen the paper.

    Advancements:

    This is a useful catalogue of potential DNA binding proteins of interest, beyond just CTCF. Some known TFs are there, but also new ones are found.

    Audience:

    Basic research mainly, with particular focus on chromatin conformation and TF binding fields.

    My expertise:

    ML/AI methods in genomics, TF binding models, epigenetics and 3D chromatin interactions.

  3. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see my points below). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the following list.

    Also, I encourage the authors to integrate the current presentation of the data with other (published) data about chromatin architecture, to make more robust the claims and go deeper into the biological implications of the current work. Se my list below.

    It follows a specific list of relevant points to be addressed:

    Specific points:

    1. Introduction, line 95: CTCF appears two times, it seems redundant;
    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?
    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS;
    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details;
    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.
    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?
    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?
    8. Do the authors think that the identified DBPs could work in that way as well?
    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?
    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Significance

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, chromatin organization is an important topic in the context of a constantly expanding research field. Therefore, the work is timely and could be useful for the community. The paper appears overall well written and the figures look clear and of good quality. Nevertheless, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see list of specific points). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the above reported points.

  4. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    The study by Osato and Hamada aims at computationally identifying a set of novel putative insulator-associated DNA binding proteins (DBPs) via estimation of their contribution to the expression of genes in the same chromosome region of their binding sites (+- 1Mbp from TSS). To achieve this, the authors leverage a deep learning architecture already published via which ChIP-seq peaks of DBPs in the TSS of a given gene are used to predict its expression level in four human cell lines.

    Building on this, the authors used another tool called DeepLIFT to evaluate the weight of each DBP binding site on the final gene expression value. Hence they made the assumption that if a given DBP had an insulator function they could restrict the prediction of the gene's expression to the region included between pairs of that DBP binding sites, and evaluate the pair's motif directionality bias in the distribution of weights. They exemplify their approach's validity by the fact that they can predict the known directionality bias of CTCF/cohesin-bound sites as the highest of the lot, with the F-R orientation of the pairs the most enriched, recapitulating what already known in literature: i.e., that F-R chromatin interaction peaks are the most enriched. In addition, they find several new DBPs showing significant directionality bias; hence they could be candidates for insulation activity. They then provide correlation between these putative insulator binding sites and sites of transition between euchromatin and heterochromatin by independently using histone mark and gene expression datasets. This, of course, is not surprising because (a) there is insulation between regions with heterotypic chromatin identities, and (b) it was already known from the first papers describing insulated chromatin domains that their boundaries were well-enriched for active transcription and transcriptional regulators (e.g., Dixon et al, Nature 2012).

    Finally, they use chromatin interaction (looping) sites to check the overlap between CTCF and all other DBPs and define a subset of putative insulator DBPs not overlapping CTCF peaks, suggesting potentially new insulatory mechanisms. These factors were all known transcriptional activators, but this part of the findings carry most of the novelty in the work and have the potential of opening up new directions for research in chromatin organization.

    Overall, the methodology applied here is adequate, clear, and reproducible. The major issue, in our view, is that the entire manuscript's findings relies on the usage of deepLIFT, a tool which was not benchmarked previously or by the current study. In fact, deepLIFT is public as regards its code, and also appears as a preprint from 2017 on biorXiv and published in the Proceedings of Machine Learning Research conference. Also, this key tool was developed by the Kundaje lab (who produce high quality alogrithms), and not by the authors. Therefore, the manuscript is predominantly based on the execution of existing workflows to publicly-available data. This does not take anything away from the interesting question posed here, but at the same time does not provide the community with any new algorithm/workflow.

    Finally, although I appreciate that the authors are purely computational and have likely no capacity for experimental validation of their claims of new DBPs having insulator roles, I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning. Using this kind of data, effects on gene expression can at least be tested in regard to the authors' predictions. Moreover, in terms of validation, Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As secondary issues, we would point out that:

    • The suggested alternative transcripts function, also highlighted in the manuscript;s abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.
    • Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.
    • Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Significance

    The scientific novelty of the work lies primarily in the identification of a set of DBPs that are proposed to confer insulator activity genome-wide. This has been long sought after in human data (whilst it is well understood and defined in Drosophila). The authors produce a quantitative ranking of the putative insulation effect of these DBPs and, most importantly, go on to identify a smaller subset that are apparently non-overlapping with anchors of CTCF-cohesin loop anchors; the presence of strong motif orientation biases in many DBPs can also be of broad interest, especially those that cannot be trivially ascribable to the loop extrusion process.

    However, although these findings open the way for speculation on multiple insulation mechanisms via proteins with multiple regulatory functions, the manuscript provide no experimental or computational means to test the proposed roles of these DBPs - and, as such, this limits the potential impact of the work and mostly targets researchers in the field of genome organization that can test these findings. Having said this, if validated, this work can significantly broaden our understanding of how chromatin is organized in 3D nuclear space.

    I typically identify myself to the authors: A. Papantonis, expertise in 3D genome architecture, chromatin biology, and genomics/bioinformatics.

  5. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Reply to the Reviewers

    I would like to thank the reviewers for their comments and interest in the manuscript and the study.

    Reviewer #1

    1. I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning.

    The directional positioning of CTCF-binding sites at chromatin interaction sites was analyzed by CRISPR experiment (Guo Y et al. Cell 2015). We found that the machine learning and statistical analysis showed the same directional bias of CTCF-binding motif sequence and RAD21-binding motif sequence at chromatin interaction sites as the experimental analysis of Guo Y et al. (lines 229-253, Figure 3b, c, d and Table 1). Since CTCF is involved in different biological functions (Braccioli L et al. Essays Biochem. 2019 ResearchGate webpage), the directional bias of binding sites may be reduced in all binding sites including those at chromatin interaction sites (lines 68-73). In our study, we investigated the DNA-binding sites of proteins using the ChIP-seq data of DNA-binding proteins and DNase-seq data. We also confirmed that the DNA-binding sites of SMC3 and RAD21, which tend to be found in chromatin loops with CTCF, also showed the same directional bias as CTCF by the computational analysis.

    __2. Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. __

    Following the reviewer's advice, I performed the same analysis with the DNA-binding sites that do no overlap with the DNA-binding sites of CTCF and cohesin (RAD21 and SMC3) (Fig. 6 and Supplementary Fig. 4). The result showed the same tendency in the distribution of DNA-binding sites. The height of a peak on the graph became lower for some DNA-binding proteins after removing the DNA-binding sites that overlapped with those of CTCF and cohesin. I have added the following sentence on lines 435 and 829: For the insulator-associated DBPs other than CTCF, RAD21, and SMC3, the DNA-binding sites that do not overlap with those of CTCF, RND21, and SMC3 were used to examine their distribution around interaction sites.

    3. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As suggested by the reviewer, I have added the insulator scores and boundary sites from the 4D nucleome data portal as tracks in the UCSC genome browser. The insulator scores seem to correspond to some extent to the H3K27me3 histone marks from ChIP-seq (Fig. 4a and Supplementary Fig. 3). We found that the DNA-binding sites of the insulator-associated DBPs were statistically overrepresented in the 5 kb boundary sites more than other DBPs (Fig. 4d). The direction of DNA-binding sites on the genome can be shown with different colors (e.g. red and green), but the directionality of insulator-associated DNA-binding sites is their overall tendency, and it may be difficult to notice the directionality from each binding site because the directionality may be weaker than that of CTCF, RAD21, and SMC3 as shown in Table 1 and Supplementary Table 2. We also observed the directional biases of CTCF, RAD21, and SMC3 by using Micro-C chromatin interaction data as we estimated, but the directionality was more apparent to distinguish the differences between the four directions of FR, RF, FF, and RR using CTCF-mediated ChIA-pet chromatin interaction data (lines 287 and 288).

     I found that the CTCF binding sites examined by a wet experiment in the previous study may not always overlap with the boundary sites of chromatin interactions from Micro-C assay (Guo Y et al. *Cell* 2015). The chromatin interaction data do not include all interactions due to the high sequencing cost of the assay, and include less long-range interactions due to distance bias. The number of the boundary sites may be smaller than that of CTCF binding sites acting as insulators and/or some of the CTCF binding sites may not be locate in the boundary sites. It may be difficult for the boundary location algorithm to identify a short boundary location. Due to the limitations of the chromatin interaction data, I planned to search for insulator-associated DNA-binding proteins without using chromatin interaction data in this study.

 I discussed other causes in lines 614-622: Another reason for the difference may be that boundary sites are more closely associated with topologically associated domains (TADs) of chromosome than are insulator sites. Boundary sites are regions identified based on the separation of numerous chromatin interactions. On the other hand, we found that the multiple DNA-binding sites of insulator-associated DNA-binding proteins were located close to each other at insulator sites and were associated with distinct nested and focal chromatin interactions, as reported by Micro-C assay. These interactions may be transient and relatively weak, such as tissue/cell type, conditional or lineage-specific interactions.

 Furthermore, I have added the statistical summary of the analysis in lines 372-395 as follows: Overall, among 20,837 DNA-binding sites of the 97 insulator-associated proteins found at insulator sites identified by H3K27me3 histone modification marks (type 1 insulator sites), 1,315 (6%) overlapped with 264 of 17,126 5kb long boundary sites, and 6,137 (29%) overlapped with 784 of 17,126 25kb long boundary sites in HFF cells. Among 5,205 DNA-binding sites of the 97 insulator-associated DNA-binding proteins found at insulator sites identified by H3K27me3 histone modification marks and transcribed regions (type 2 insulator sites), 383 (7%) overlapped with 74 of 17,126 5-kb long boundary sites, 1,901 (37%) overlapped with 306 of 17,126 25-kb long boundary sites. Although CTCF-binding sites separate active and repressive domains, the limited number of DNA-binding sites of insulator-associated proteins found at type 1 and 2 insulator sites overlapped boundary sites identified by chromatin interaction data. Furthermore, by analyzing the regulatory regions of genes, the DNA-binding sites of the 97 insulator-associated DNA-binding proteins were found (1) at the type 1 insulator sites (based on H3K27me3 marks) in the regulatory regions of 3,170 genes, (2) at the type 2 insulator sites (based on H3K27me3 marks and gene expression levels) in the regulatory regions of 1,044 genes, and (3) at insulator sites as boundary sites identified by chromatin interaction data in the regulatory regions of 6,275 genes. The boundary sites showed the highest number of overlaps with the DNA-binding sites. Comparing the insulator sites identified by (1) and (3), 1,212 (38%) genes have both types of insulator sites. Comparing the insulator sites between (2) and (3), 389 (37%) genes have both types of insulator sites. From the comparison of insulator and boundary sites, we found that (1) or (2) types of insulator sites overlapped or were close to boundary sites identified by chromatin interaction data.

    4. The suggested alternative transcripts function, also highlighted in the manuscripts abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.

    According to the reviewer's comment, I performed the genome-wide analysis of alternative transcripts where the DNA-binding sites of insulator-associated proteins are located near splicing sites. The DNA-binding sites of insulator-associated DNA-binding proteins were found within 200 bp centered on splice sites more significantly than the other DNA-binding proteins (Fig. 4e and Table 2). I have added the following sentences on lines 405 - 412: We performed the statistical test to estimate the enrichment of insulator-associated DNA-binding sites compared to the other DNA-binding proteins, and found that the insulator-associated DNA-binding sites were significantly more abundant at splice sites than the DNA-binding sites of the other proteins (Fig 4e and Table 2; Mann‒Whitney U test, p value 5. Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.

    I believe that the Figure 1 would help researchers in other fields who are not familiar with biological phenomena and functions to understand the study. More explanation has been included in the Figures and legends of Figs. 4 and 5 to help readers outside the immediate research field understand the figures.

    6. Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Reviewer #2's comments would be related to this comment. I have introduced a more detailed explanation of the method in the Results section, as shown in the responses to Reviewer #2's comments.

    Reviewer #2

    1. Introduction, line 95: CTCF appears two times, it seems redundant.

    On lines 91-93, I deleted the latter CTCF from the sentence "We examine the directional bias of DNA-binding sites of CTCF and insulator-associated DBPs, including those of known DBPs such as RAD21 and SMC3".

    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?

    Although CTCF is known to be the main insulator protein in vertebrates, we found that 97 DNA-binding proteins including CTCF and cohesin are associated with insulator sites by modifying and developing a machine learning method to search for insulator-associated DNA-binding proteins. Most of the insulator-associated DNA-binding proteins showed the directional bias of DNA-binding motifs, suggesting that the directional bias is associated with the insulator.

     I have added the sentence in lines 96-99 as follows: Furthermore, statistical testing the contribution scores between the directional and non-directional DNA-binding sites of insulator-associated DBPs revealed that the directional sites contributed more significantly to the prediction of gene expression levels than the non-directional sites. I have revised the statement in lines 101-110 as follows: To validate these findings, we demonstrate that the DNA-binding sites of the identified insulator-associated DBPs are located within potential insulator sites, and some of the DNA-binding sites in the insulator site are found without the nearby DNA-binding sites of CTCF and cohesin. Homologous and heterologous insulator-insulator pairing interactions are orientation-dependent, as suggested by the insulator-pairing model based on experimental analysis in flies. Our method and analyses contribute to the identification of insulator- and chromatin-associated DNA-binding sites that influence EPIs and reveal novel functional roles and molecular mechanisms of DBPs associated with transcriptional condensation, phase separation and transcriptional regulation.

    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS.

    On lines 121-124, to explain the procedure for the SNP of an eQTL, I have added the sentence in the Methods: "If a DNA-binding site was located within a 100-bp region around a single-nucleotide polymorphism (SNP) of an eQTL, we assumed that the DNA-binding proteins regulated the expression of the transcript corresponding to the eQTL".

    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details.

    On line 119, I have included the explanation of the eQTL dataset of GTEx v8 as follows: " The eQTL data were derived from the GTEx v8 dataset, after quality control, consisting of 838 donors and 17,382 samples from 52 tissues and two cell lines". On lines 681 and 865, I have added the filename of the eQTL data "(GTEx_Analysis_v8_eQTL.tar)".

    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.

    The reviewer would mention Figure 2, not Figure 1. If so, the matrices in panels a and b in Figure 2 are equivalent. I have shown it in the figure: The same figure in panel a is rotated 90 degrees to the right. The green boxes in the matrix show the regions with the ChIP-seq peak of a DNA-binding protein overlapping with a SNP of an eQTL. I used eQTL data to associate a gene with a ChIP-seq peak that was more than 2 kb upstream and 1 kb downstream of a transcriptional start site of a gene. For each gene, the matrix was produced and the gene expression levels in cells were learned and predicted using the deep learning method. I have added the following sentences to explain the method in lines 133 - 139: Through the training, the tool learned to select the binding sites of DNA-binding proteins from ChIP-seq assays that were suitable for predicting gene expression levels in the cell types. The binding sites of a DNA-binding protein tend to be observed in common across multiple cell and tissue types. Therefore, ChIP-seq data and eQTL data in different cell and tissue types were used as input data for learning, and then the tool selected the data suitable for predicting gene expression levels in the cell types, even if the data were not obtained from the same cell types.

    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?

    As suggested by the reviewer, to help readers understand the observation, I have added Supplementary Fig. S4c to show the distribution of DNA-binding sites of "CTCF, RAD21, and SMC3" and "BACH2, FOS, ATF3, NFE2, and MAFK" around chromatin interaction sites. I have modified the following sentence to indicate the figure on line 501: Although a DNA-binding-site distribution pattern around chromatin interaction sites similar to those of CTCF, RAD21, and SMC3 was observed for DBPs such as BACH2, FOS, ATF3, NFE2, and MAFK, less than 1% of the DNA-binding sites of the latter set of DBPs colocalized with CTCF, RAD21, or SMC3 in a single bin (Fig. S4c).

     In Aljahani A et al. *Nature Communications* 2022, we find that depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Together, our data show that loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression. Goel VY et al. *Nature Genetics* 2023 mentioned in the abstract: Microcompartments frequently connect enhancers and promoters and though loss of loop extrusion and inhibition of transcription disrupts some microcompartments, most are largely unaffected. These results suggested that chromatin loops can be driven by other DBPs independent of the known CTCF/Cohesin.

I added the following sentence on lines 569-577: The depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. Furthermore, the loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression.

 FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates (Ji D et al. *Molecular Cell* 2024). CTCF have also found to form transcriptional condensate and phase separation (Lee R et al. *Nucleic acids research* 2022). FOS was found to be an insulator-associated DNA-binding protein in this study and is potentially involved in chromatin remodeling, transcription condensation, and phase separation with the other factors such as BACH2, ATF3, NFE2 and MAFK. I have added the following sentence on line 556: FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates.

    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?

    Goel VY et al. Nature Genetics 2023 identified highly nested and focal interactions through region capture Micro-C, which resemble fine-scale compartmental interactions and are termed microcompartments. In the section titled "Most microcompartments are robust to loss of loop extrusion," the researchers noted that a small proportion of interactions between CTCF and cohesin-bound sites exhibited significant reductions in strength when cohesin was depleted. In contrast, the majority of microcompartmental interactions remained largely unchanged under cohesin depletion. Our findings indicate that most P-P and E-P interactions, aside from a few CTCF and cohesin-bound enhancers and promoters, are likely facilitated by a compartmentalization mechanism that differs from loop extrusion. We suggest that nested, multiway, and focal microcompartments correspond to small, discrete A-compartments that arise through a compartmentalization process, potentially influenced by factors upstream of RNA Pol II initiation, such as transcription factors, co-factors, or active chromatin states. It follows that if active chromatin regions at microcompartment anchors exhibit selective "stickiness" with one another, they will tend to co-segregate, leading to the development of nested, focal interactions. This microphase separation, driven by preferential interactions among active loci within a block copolymer, may account for the striking interaction patterns we observe.

     The authors of the paper proposed several mechanisms potentially involved in microcompartments. These mechanisms may be involved in looping with insulator function. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently (Hsieh TS et al. *Nature Genetics* 2022). Among the identified insulator-associated DNA-binding proteins, Maz and MyoD1 form loops without CTCF (Xiao T et al. *Proc Natl Acad Sci USA* 2021 ; Ortabozkoyun H et al. *Nature genetics* 2022 ; Wang R et al. *Nature communications* 2022). I have added the following sentences on lines 571-575: Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. I have included the following explanation on lines 582-584: Maz and MyoD1 among the identified insulator-associated DNA-binding proteins form loops without CTCF.

 As for the directionality of CTCF, if chromatin loop anchors have some structural conformation, as shown in the paper entitled "The structural basis for cohesin-CTCF-anchored loops" (Li Y et al. *Nature* 2020), directional DNA binding would occur similarly to CTCF binding sites. Moreover, cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops (Davidson IF et al. *Nature Reviews Molecular Cell Biology* 2021). Regarding loop extrusion, the 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions (Guerin TM et al. *EMBO Journal* 2024). I have added the following sentences on lines 543-547: Cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops. I have included the following sentences on lines 577-582: The 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions.

 Another model for the regulation of gene expression by insulators is the boundary-pairing (insulator-pairing) model (Bing X et al. *Elife* 2024) (Ke W et al. *Elife* 2024) (Fujioka M et al. *PLoS Genetics* 2016). Molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies. Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent. I have summarized the model on lines 559-567: Other types of chromatin regulation are also expected to be related to the structural interactions of molecules. As the boundary-pairing (insulator-pairing) model, molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies (Fig. 7). Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent.

    8. Do the authors think that the identified DBPs could work in that way as well?

    The boundary-pairing (insulator-pairing) model would be applied to the insulator-associated DNA-binding proteins other than CTCF and cohesin that are involved in the loop extrusion mechanism (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016).

     Liquid-liquid phase separation was shown to occur through CTCF-mediated chromatin loops and to act as an insulator (Lee, R et al. *Nucleic Acids Research* 2022). Among the identified insulator-associated DNA-binding proteins, CEBPA has been found to form hubs that colocalize with transcriptional co-activators in a native cell context, which is associated with transcriptional condensate and phase separation (Christou-Kent M et al. *Cell Reports* 2023). The proposed microcompartment mechanisms are also associated with phase separation. Thus, the same or similar mechanisms are potentially associated with the insulator function of the identified DNA-binding proteins. I have included the following information on line 554: CEBPA in the identified insulator-associated DNA-binding proteins was also reported to be involved in transcriptional condensates and phase separation.

    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?

    Snead WT et al. Molecular Cell 2019 mentioned that protein post-transcriptional modifications (PTMs) facilitate the control of molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin (Tang X et al. Nature Communications 2024). I found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Supplementary Fig. 2d). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation by PTMs. I have added the following explanation on lines 584-590: Furthermore, protein post-transcriptional modifications (PTMs) facilitate control over the molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin. We found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Fig. 4f and Supplementary Fig. 3c). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation through PTMs.

    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Structural molecular model of cohesin-CTCF-anchored loops has been published by Li Y et al. Nature 2020. The structural conformation of CTCF and cohesin in the loops would be the cause of the directional bias of CTCF binding sites, which I mentioned in lines 539 - 543 as follows: These results suggest that the directional bias of DNA-binding sites of insulator-associated DBPs may be involved in insulator function and chromatin regulation through structural interactions among DBPs, other proteins, DNAs, and RNAs. For example, the N-terminal amino acids of CTCF have been shown to interact with RAD21 in chromatin loops.

     To investigate the principles underlying the architectural functions of insulator-insulator pairing interactions, two insulators, Homie and Nhomie, flanking the *Drosophila even skipped *locus were analyzed. Pairing interactions between the transgene Homie and the eve locus are directional. The head-to-head pairing between the transgene and endogenous Homie matches the pattern of activation (Fujioka M et al. *PLoS Genetics* 2016).

    Reviewer #3

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.

    When a protein complex binds to DNA, one protein of the complex binds to the DNA directory, and the other proteins may not bind to DNA. However, the DNA motif sequence bound by the protein may be registered as the DNA-binding motif of all the proteins in the complex. The molecular structure of the complex of CTCF and Cohesin showed that both CTCF and Cohesin bind to DNA (Li Y et al. Nature 2020). I think there is a possibility that if the molecular structure of a protein complex becomes available, the previous recognition of the DNA-binding ability of a protein may be changed. Therefore, I searched the Pfam database for 99 insulator-associated DNA-binding proteins identified in this study. I found that 97 are registered as DNA-binding proteins and/or have a known DNA-binding domain, and EP300 and SIN3A do not directory bind to DNA, which was also checked by Google search. I have added the following explanation in line 257 to indicate direct and indirect DNA-binding proteins: Among 99 insulator-associated DBPs, EP300 and SIN3A do not directory interact with DNA, and thus 97 insulator-associated DBPs directory bind to DNA. I have updated the sentence in line 20 of the Abstract as follows: We discovered 97 directional and minor nondirectional motifs in human fibroblast cells that corresponded to 23 DBPs related to insulator function, CTCF, and/or other types of chromosomal transcriptional regulation reported in previous studies.

    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.

    As the reviewer mentioned, I recognize enhancers are relatively small regions. In the paper, I intended to examine further upstream and downstream of promoter regions where enhancers are found. Therefore, I have modified the sentence in lines 929 - 931 of the Fig. 2 legend as follows: Enhancer-gene regulatory interaction regions consist of 200 bins of 10 kbp between -1 Mbp and 1 Mbp region from TSS, not including promoter.

    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.

    Following the reviewer's advice, I have added the ChIP-seq data of H3K9me3 as a truck of the UCSC Genome Browser. The distribution of H3K9me3 signal was different from that of H3K27me3 in some regions. I also found the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions and took some screenshots of the UCSC Genome Browser of the regions around the sites in Supplementary Fig. 3b. I have modified the following sentence on lines 974 - 976 in the legend of Fig. 4: a Distribution of histone modification marks H3K27me3 (green color) and H3K9me3 (turquoise color) and transcript levels (pink color) in upstream and downstream regions of a potential insulator site (light orange color). I have also added the following result on lines 356 - 360: The same analysis was performed using H3K9me3 marks, instead of H3K27me3 (Fig. S3b). We found that the distribution of H3K9me3 signal was different from that of H3K27me3 in some regions, and discovered the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions (Fig. S3b).

    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    The resolution of the Micro-C assay is considered to be 100 bp and above, as the human nucleome core particle contains 145 bp (and 193 bp with linker) of DNA. However, internucleosomal DNA is cleaved by endonuclease into fragments of multiples of 10 nucleotides (Pospelov VA et al. *Nucleic Acids Research *1979). Highly nested focal interactions were observed (Goel VY et al. *Nature Genetics *2023). Base pair resolution was reported using Micro Capture-C (Hua P et al. Nature 2021). Sub-kilobase (20 bp resolution) chromatin topology was reported using an MNase-based chromosome conformation capture (3C) approach (Aljahani A et al. *Nature Communications *2022). On the other hand, Hi-C data was analyzed at 1 kb resolution. (Gu H et al. bioRxiv 2021). If the resolution of Micro-C interactions is at best at 1 kb, the binding sites of a DNA-binding protein will not show a peak around the center of the genomic locations of interaction edges. Each panel shows the number of binding sites of a specific DNA-binding protein at a specific distance from the midpoint of all chromatin interaction edges. I have modified and added the following sentences in lines 593-597: High-resolution chromatin interaction data from a Micro-C assay indicated that most of the predicted insulator-associated DBPs showed DNA-binding-site distribution peaks around chromatin interaction sites, suggesting that these DBPs are involved in chromatin interactions and that the chromatin interaction data has a high degree of resolution. Base pair resolution was reported using Micro Capture-C.

    Minor Comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., ____https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2____or ____https://pubmed.ncbi.nlm.nih.gov/37486787____/). The authors should discuss how that would impact their results.

    The directional bias of CTCF binding sites was identified by ChIA-pet interactions of CTCF binding sites. The analysis of the contribution scores of DNA-binding sites of proteins considering the binding sites of CTCF as an insulator showed the same tendency of directional bias of CTCF binding sites. In the analysis, to remove the false-positive prediction of DNA-binding sites, I used the binding sites that overlapped with a ChIP-seq peak of the DNA-binding protein. This result suggests that the DNA-binding sites of CTCF obtained by the current analysis have sufficient quality. Therefore, if the accuracy of prediction of DNA-binding sites is improved, although the number of DNA-binding sites may be different, the overall tendency of the directionality of DNA-binding sites will not change and the results of this study will not change significantly.

     As for the first reference in the reviewer's comment, chromatin interaction data from Micro-C assay does not include all chromatin interactions in a cell or tissue, because it is expensive to cover all interactions. Therefore, it would be difficult to predict all chromatin interactions based on machine learning. As for the second reference in the reviewer's comment, pioneer factors such as FOXA are known to bind to closed chromatin regions, but transcription factors and DNA-binding proteins involved in chromatin interactions and insulators generally bind to open chromatin regions. The search for the DNA-binding motifs is not required in closed chromatin regions.

    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    In the DeepLIFT paper, the authors explain that DeepLIFT is a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input (Shrikumar A et al. ICML 2017). DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

     Truly explainable AI would be able to find cause and reason, and to make choices and decisions like humans. DeepLIFT does not perform causal inferences. I did not use the term "Explainable AI" in our manuscript, but I briefly explained it in Discussion. I have added the following explanation in lines 623-628: AI (Artificial Intelligence) is considered as a black box, since the reason and cause of prediction are difficult to know. To solve this issue, tools and methods have been developed to know the reason and cause. These technologies are called Explainable AI. DeepLIFT is considered to be a tool for Explainable AI. However, DeepLIFT does not answer the reason and cause for a prediction. It calculates scores representing the contribution of the input data to the prediction.

 Furthermore, to improve the readability of the manuscript, I have included the following explanation in lines 159-165: we computed DeepLIFT scores of the input data (i.e., each binding site of the ChIP-seq data of DNA-binding proteins) in the deep leaning analysis on gene expression levels. DeepLIFT compares the importance of each input for predicting gene expression levels to its 'reference or background level' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.
  6. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary:

    Osato and Hamada propose a systematic approach to identify DNA binding proteins that display directional binding. They used a modified Deep Learning method (DEcode) to investigate binding profiles of 1356 DBP from GTRD database at promoters (30 of 100bp bins around TSS) and enhancers (200 bins of 10Kb around eSNPs) and use this to predict expression of 25,071 genes in Fibroblasts, Monocytes, HMEC and NPC. This method achieves a good prediction power (Spearman correlation between predicted and actual expression of 0.74). They then use PIQ, and overlap predicted binding sites with actual ChIP-seq data to investigate the motifs of TFs that are controlling gene expression. They find 99 insulator proteins showing either a specific directional bias or minor non-directional bias, corresponding to 23 DBP previously reported to have insulator function. Of the 23 proteins they identify as regulating enhancer promoter interactions, 13 are associated with CTCF. They also show that there are significantly more insulator proteins binding sites at borders of polycomb domains, transcriptionally active or boundary regions based on chromatin interactions than other proteins.

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.
    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.
    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.
    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    Minor comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.
    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    Referee Cross-Commenting

    I would like to mention that I agree with the comments of reviewers 1 and 2.

    Significance

    General assessment:

    This is the first study to my knowledge that attempts to use Deep Learning to identify insulators and directional biases in binding. One of the limitations is that no additional methods were used to show that these DBP have directional binding bias. It is not necessarily to employ additional methods, but it would definitely strengthen the paper.

    Advancements:

    This is a useful catalogue of potential DNA binding proteins of interest, beyond just CTCF. Some known TFs are there, but also new ones are found.

    Audience:

    Basic research mainly, with particular focus on chromatin conformation and TF binding fields.

    My expertise:

    ML/AI methods in genomics, TF binding models, epigenetics and 3D chromatin interactions.

  7. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see my points below). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the following list.

    Also, I encourage the authors to integrate the current presentation of the data with other (published) data about chromatin architecture, to make more robust the claims and go deeper into the biological implications of the current work. Se my list below.

    It follows a specific list of relevant points to be addressed:

    Specific points:

    1. Introduction, line 95: CTCF appears two times, it seems redundant;
    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?
    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS;
    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details;
    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.
    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?
    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?
    8. Do the authors think that the identified DBPs could work in that way as well?
    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?
    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Significance

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, chromatin organization is an important topic in the context of a constantly expanding research field. Therefore, the work is timely and could be useful for the community. The paper appears overall well written and the figures look clear and of good quality. Nevertheless, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see list of specific points). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the above reported points.

  8. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    The study by Osato and Hamada aims at computationally identifying a set of novel putative insulator-associated DNA binding proteins (DBPs) via estimation of their contribution to the expression of genes in the same chromosome region of their binding sites (+- 1Mbp from TSS). To achieve this, the authors leverage a deep learning architecture already published via which ChIP-seq peaks of DBPs in the TSS of a given gene are used to predict its expression level in four human cell lines.

    Building on this, the authors used another tool called DeepLIFT to evaluate the weight of each DBP binding site on the final gene expression value. Hence they made the assumption that if a given DBP had an insulator function they could restrict the prediction of the gene's expression to the region included between pairs of that DBP binding sites, and evaluate the pair's motif directionality bias in the distribution of weights. They exemplify their approach's validity by the fact that they can predict the known directionality bias of CTCF/cohesin-bound sites as the highest of the lot, with the F-R orientation of the pairs the most enriched, recapitulating what already known in literature: i.e., that F-R chromatin interaction peaks are the most enriched. In addition, they find several new DBPs showing significant directionality bias; hence they could be candidates for insulation activity. They then provide correlation between these putative insulator binding sites and sites of transition between euchromatin and heterochromatin by independently using histone mark and gene expression datasets. This, of course, is not surprising because (a) there is insulation between regions with heterotypic chromatin identities, and (b) it was already known from the first papers describing insulated chromatin domains that their boundaries were well-enriched for active transcription and transcriptional regulators (e.g., Dixon et al, Nature 2012).

    Finally, they use chromatin interaction (looping) sites to check the overlap between CTCF and all other DBPs and define a subset of putative insulator DBPs not overlapping CTCF peaks, suggesting potentially new insulatory mechanisms. These factors were all known transcriptional activators, but this part of the findings carry most of the novelty in the work and have the potential of opening up new directions for research in chromatin organization.

    Overall, the methodology applied here is adequate, clear, and reproducible. The major issue, in our view, is that the entire manuscript's findings relies on the usage of deepLIFT, a tool which was not benchmarked previously or by the current study. In fact, deepLIFT is public as regards its code, and also appears as a preprint from 2017 on biorXiv and published in the Proceedings of Machine Learning Research conference. Also, this key tool was developed by the Kundaje lab (who produce high quality alogrithms), and not by the authors. Therefore, the manuscript is predominantly based on the execution of existing workflows to publicly-available data. This does not take anything away from the interesting question posed here, but at the same time does not provide the community with any new algorithm/workflow.

    Finally, although I appreciate that the authors are purely computational and have likely no capacity for experimental validation of their claims of new DBPs having insulator roles, I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning. Using this kind of data, effects on gene expression can at least be tested in regard to the authors' predictions. Moreover, in terms of validation, Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As secondary issues, we would point out that:

    • The suggested alternative transcripts function, also highlighted in the manuscript;s abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.
    • Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.
    • Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Significance

    The scientific novelty of the work lies primarily in the identification of a set of DBPs that are proposed to confer insulator activity genome-wide. This has been long sought after in human data (whilst it is well understood and defined in Drosophila). The authors produce a quantitative ranking of the putative insulation effect of these DBPs and, most importantly, go on to identify a smaller subset that are apparently non-overlapping with anchors of CTCF-cohesin loop anchors; the presence of strong motif orientation biases in many DBPs can also be of broad interest, especially those that cannot be trivially ascribable to the loop extrusion process.

    However, although these findings open the way for speculation on multiple insulation mechanisms via proteins with multiple regulatory functions, the manuscript provide no experimental or computational means to test the proposed roles of these DBPs - and, as such, this limits the potential impact of the work and mostly targets researchers in the field of genome organization that can test these findings. Having said this, if validated, this work can significantly broaden our understanding of how chromatin is organized in 3D nuclear space.

    I typically identify myself to the authors: A. Papantonis, expertise in 3D genome architecture, chromatin biology, and genomics/bioinformatics.

  9. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Reply to the Reviewers

    I would like to thank the reviewers for their comments and interest in the manuscript and the study.

    Referee #1

    1. I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning.

    Response: The directional positioning of CTCF-binding sites at chromatin interaction sites was analyzed by CRISPR experiment (Guo Y et al. Cell 2015). We found that the machine learning and statistical analysis showed the same directional bias of the CTCF-binding motif sequence at chromatin interaction sites as the experimental analysis of Guo Y et al. (lines 229-245, Figure 3b, c, d and Table 1). Since CTCF is involved in different biological functions (Braccioli L et al. Essays Biochem. 2019 ResearchGate webpage), the directional bias of binding sites may be reduced in all binding sites including those at chromatin interaction sites (lines 68-73). In our study, we investigated the DNA-binding sites of proteins using the ChIP-seq data of DNA-binding proteins and DNase-seq data. We also confirmed that the DNA-binding sites of SMC3 and RAD21, which tend to be found in chromatin loops with CTCF, also showed the same directional bias as CTCF by the computational analysis.

    1. Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure.

    Response: Following the reviewer's advice, I performed the same analysis with the DNA-binding sites that do no overlap with the DNA-binding sites of CTCF and cohesin (RAD21 and SMC3) (Fig. 6 and Supplementary Fig. 4). The result showed the same tendency in the distribution of DNA-binding sites. The height of a peak on the graph became lower for some DNA-binding proteins after removing the DNA-binding sites that overlapped with those of CTCF and cohesin. I have added the following sentence on lines 427 and 817: For the insulator-associated DBPs other than CTCF, RAD21, and SMC3, the DNA-binding sites that do not overlap with those of CTCF, RND21, and SMC3 were used to examine their distribution around interaction sites.

    1. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    Response: As suggested by the reviewer, I have added the insulator scores and boundary sites from the 4D nucleome data portal as tracks in the UCSC genome browser. The insulator scores seem to correspond to some extent to the H3K27me3 histone marks from ChIP-seq (Fig. 4a and Supplementary Fig. 3). The direction of DNA-binding sites on the genome can be shown with different colors (e.g. red and green), but the directionality of insulator-associated DNA-binding sites is their overall tendency, and it may be difficult to notice the directionality from each binding site because the directionality may be weaker than that of CTCF, RAD21, and SMC3 as shown in Table 1 and Supplementary Table 2.

    I found that the CTCF binding sites examined by a wet experiment in the previous study may not always overlap with the boundary sites of chromatin interactions from Micro-C assay (Guo Y et al. Cell 2015). The chromatin interaction data do not include all interactions due to the high sequencing cost of the assay. The number of the boundary sites may be smaller than that of CTCF binding sites acting as insulators and/or some of the CTCF binding sites may not be locate in the boundary sites. It may be difficult for the boundary location algorithm to identify a short boundary location. Due to the limitations of the chromatin interaction data, I planned to search for insulator-associated DNA-binding proteins without using chromatin interaction data in this study. I have added the statistical summary of the analysis in lines 364-387 as follows: Overall, among 20,837 DNA-binding sites of the 97 insulator-associated proteins found at insulator sites identified by H3K27me3 histone modification marks (type 1 insulator sites), 1,315 (6%) overlapped with 264 of 17,126 5kb long boundary sites, and 6,137 (29%) overlapped with 784 of 17,126 25kb long boundary sites in HFF cells. Among 5,205 DNA-binding sites of the 97 insulator-associated DNA-binding proteins found at insulator sites identified by H3K27me3 histone modification marks and transcribed regions (type 2 insulator sites), 383 (7%) overlapped with 74 of 17,126 5-kb long boundary sites, 1,901 (37%) overlapped with 306 of 17,126 25-kb long boundary sites. Although CTCF-binding sites separate active and repressive domains, the limited number of DNA-binding sites of insulator-associated proteins found at type 1 and 2 insulator sites overlapped boundary sites identified by chromatin interaction data. Furthermore, by analyzing the regulatory regions of genes, the DNA-binding sites of the 97 insulator-associated DNA-binding proteins were found (1) at the type 1 insulator sites (based on H3K27me3 marks) in the regulatory regions of 3,170 genes, (2) at the type 2 insulator sites (based on H3K27me3 marks and gene expression levels) in the regulatory regions of 1,044 genes, and (3) at insulator sites as boundary sites identified by chromatin interaction data in the regulatory regions of 6,275 genes. The boundary sites showed the highest number of overlaps with the DNA-binding sites. Comparing the insulator sites identified by (1) and (3), 1,212 (38%) genes have both types of insulator sites. Comparing the insulator sites between (2) and (3), 389 (37%) genes have both types of insulator sites. From the comparison of insulator and boundary sites, we found that (1) or (2) types of insulator sites overlapped or were close to boundary sites identified by chromatin interaction data.

    1. The suggested alternative transcripts function, also highlighted in the manuscripts abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.

    Response: According to the reviewer's comment, I performed the genome-wide analysis of alternative transcripts where the DNA-binding sites of insulator-associated proteins are located near splicing sites. The DNA-binding sites of insulator-associated DNA-binding proteins were found within 200 bp centered on splice sites more significantly than the other DNA-binding proteins (Fig. 4e and Table 2). I have added the following sentences on lines 397 - 404: We performed the statistical test to estimate the enrichment of insulator-associated DNA-binding sites compared to the other DNA-binding proteins, and found that the insulator-associated DNA-binding sites were significantly more abundant at splice sites than the DNA-binding sites of the other proteins (Fig 4e and Table 2; Mann‒Whitney U test, p value 5. Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.

    Response: I believe that the Figure 1 would help researchers in other fields who are not familiar with biological phenomena and functions to understand the study. More explanation has been included in the Figures and legends of Figs. 4 and 5 to help readers outside the immediate research field understand the figures.

    1. Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Response: Reviewer #2's comments would be related to this comment. I have introduced a more detailed explanation of the method in the Results section, as shown in the responses to Reviewer #2's comments.

    Referee #2

    1. Introduction, line 95: CTCF appears two times, it seems redundant.

    Response: On lines 91-93, I deleted the latter CTCF from the sentence "and examined the directional bias of DNA-binding sites of CTCF and insulator-associated DBPs, including those of known DBPs such as RAD21 and SMC3".

    1. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?

    Response: Although CTCF is known to be the main insulator protein in vertebrates, we found that 97 DNA-binding proteins including CTCF and cohesin are associated with insulator sites by modifying and developing a machine learning method to search for insulator-associated DNA-binding proteins. Most of the insulator-associated DNA-binding proteins showed the directional bias of DNA-binding motifs, suggesting that the directional bias is associated with the insulator.

    I have added the sentence in lines 96-99 as follows: Furthermore, statistical testing the contribution scores between the directional and non-directional DNA-binding sites of insulator-associated DBPs revealed that the directional sites contributed more significantly to the prediction of gene expression levels than the non-directional sites. I have revised the statement in lines 101-110 as follows: To validate these findings, we demonstrate that the DNA-binding sites of the identified insulator-associated DBPs are located within potential insulator sites, and some of the DNA-binding sites in the insulator site are found without the nearby DNA-binding sites of CTCF and cohesin. Homologous and heterologous insulator-insulator pairing interactions are orientation-dependent, as suggested by the insulator-pairing model based on experimental analysis in flies. Our method and analyses contribute to the identification of insulator- and chromatin-associated DNA-binding sites that influence EPIs and reveal novel functional roles and molecular mechanisms of DBPs associated with transcriptional condensation, phase separation and transcriptional regulation.

    1. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS.

    Response: On lines 121-124, to explain the procedure for the SNP of an eQTL, I have added the sentence in the Methods: "If a DNA-binding site was located within a 100-bp region around a single-nucleotide polymorphism (SNP) of an eQTL, we assumed that the DNA-binding proteins regulated the expression of the transcript corresponding to the eQTL".

    1. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details.

    Response: On line 119, I have included the explanation of the eQTL dataset of GTEx v8 as follows: " The eQTL data were derived from the GTEx v8 dataset, after quality control, consisting of 838 donors and 17,382 samples from 52 tissues and two cell lines". On lines 681 and 865, I have added the filename of the eQTL data "(GTEx_Analysis_v8_eQTL.tar)".

    1. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.

    Response: The reviewer would mention Figure 2, not Figure 1. If so, the matrices in panels a and b in Figure 2 are equivalent. I have shown it in the figure: The same figure in panel a is rotated 90 degrees to the right. The green boxes in the matrix show the regions with the ChIP-seq peak of a DNA-binding protein overlapping with a SNP of an eQTL. I used eQTL data to associate a gene with a ChIP-seq peak that was more than 2 kb upstream and 1 kb downstream of a transcriptional start site of a gene. For each gene, the matrix was produced and the gene expression levels in cells were learned and predicted using the deep learning method. I have added the following sentences to explain the method in lines 133 - 139: Through the training, the tool learned to select the binding sites of DNA-binding proteins from ChIP-seq assays that were suitable for predicting gene expression levels in the cell types. The binding sites of a DNA-binding protein tend to be observed in common across multiple cell and tissue types. Therefore, ChIP-seq data and eQTL data in different cell and tissue types were used as input data for learning, and then the tool selected the data suitable for predicting gene expression levels in the cell types, even if the data were not obtained from the same cell types.

    1. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?

    Response: As suggested by the reviewer, to help readers understand the observation, I have added Supplementary Fig. S4c to show the distribution of DNA-binding sites of "CTCF, RAD21, and SMC3" and "BACH2, FOS, ATF3, NFE2, and MAFK" around chromatin interaction sites. I have modified the following sentence to indicate the figure on line 493: Although a DNA-binding-site distribution pattern around chromatin interaction sites similar to those of CTCF, RAD21, and SMC3 was observed for DBPs such as BACH2, FOS, ATF3, NFE2, and MAFK, less than 1% of the DNA-binding sites of the latter set of DBPs colocalized with CTCF, RAD21, or SMC3 in a single bin (Fig. S4c).

    In Aljahani A et al. Nature Communications 2022, we find that depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Together, our data show that loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression. Goel VY et al. Nature Genetics 2023 mentioned in the abstract: Microcompartments frequently connect enhancers and promoters and though loss of loop extrusion and inhibition of transcription disrupts some microcompartments, most are largely unaffected. These results suggested that chromatin loops can be driven by other DBPs independent of the known CTCF/Cohesin.

    FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates (Ji D et al. Molecular Cell 2024). CTCF have also found to form transcriptional condensate and phase separation (Lee R et al. Nucleic acids research 2022). FOS was found to be an insulator-associated DNA-binding protein in this study and is potentially involved in chromatin remodeling, transcription condensation, and phase separation with the other factors such as BACH2, ATF3, NFE2 and MAFK. I have added the following sentence on line 548: FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates.

    1. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?

    Response: Goel VY et al. Nature Genetics 2023 identified highly nested and focal interactions through region capture Micro-C, which resemble fine-scale compartmental interactions and are termed microcompartments. In the section titled "Most microcompartments are robust to loss of loop extrusion," the researchers noted that a small proportion of interactions between CTCF and cohesin-bound sites exhibited significant reductions in strength when cohesin was depleted. In contrast, the majority of microcompartmental interactions remained largely unchanged under cohesin depletion. Our findings indicate that most P-P and E-P interactions, aside from a few CTCF and cohesin-bound enhancers and promoters, are likely facilitated by a compartmentalization mechanism that differs from loop extrusion. We suggest that nested, multiway, and focal microcompartments correspond to small, discrete A-compartments that arise through a compartmentalization process, potentially influenced by factors upstream of RNA Pol II initiation, such as transcription factors, co-factors, or active chromatin states. It follows that if active chromatin regions at microcompartment anchors exhibit selective "stickiness" with one another, they will tend to co-segregate, leading to the development of nested, focal interactions. This microphase separation, driven by preferential interactions among active loci within a block copolymer, may account for the striking interaction patterns we observe.

    The authors of the paper proposed several mechanisms potentially involved in microcompartments. These mechanisms may be involved in looping with insulator function. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently (Hsieh TS et al. Nature Genetics 2022). Among the identified insulator-associated DNA-binding proteins, Maz and MyoD1 form loops without CTCF (Xiao T et al. Proc Natl Acad Sci USA 2021 ; Ortabozkoyun H et al. Nature genetics 2022 ; Wang R et al. Nature communications 2022). I have added the following sentences on lines 563-567: Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. I have included the following explanation on lines 574-576: Maz and MyoD1 among the identified insulator-associated DNA-binding proteins form loops without CTCF.

    As for the directionality of CTCF, if chromatin loop anchors have some structural conformation, as shown in the paper entitled "The structural basis for cohesin-CTCF-anchored loops" (Li Y et al. Nature 2020), directional DNA binding would occur similarly to CTCF binding sites. Moreover, cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops (Davidson IF et al. Nature Reviews Molecular Cell Biology 2021). Regarding loop extrusion, the 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions (Guerin TM et al. EMBO Journal 2024). I have added the following sentences on lines 535-539: Cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops. I have included the following sentences on lines 569-574: The 'loop extrusion' hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions.

    Another model for the regulation of gene expression by insulators is the boundary-pairing (insulator-pairing) model (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016). Molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies. Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent. I have summarized the model on lines 551-559: Other types of chromatin regulation are also expected to be related to the structural interactions of molecules. As the boundary-pairing (insulator-pairing) model, molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies (Fig. 7). Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent.

    1. Do the authors think that the identified DBPs could work in that way as well?

    Response: The boundary-pairing (insulator-pairing) model would be applied to the insulator-associated DNA-binding proteins other than CTCF and cohesin that are involved in the loop extrusion mechanism (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016).

    Liquid-liquid phase separation was shown to occur through CTCF-mediated chromatin loops and to act as an insulator (Lee, R et al. Nucleic Acids Research 2022). Among the identified insulator-associated DNA-binding proteins, CEBPA has been found to form hubs that colocalize with transcriptional co-activators in a native cell context, which is associated with transcriptional condensate and phase separation (Christou-Kent M et al. Cell Reports 2023). The proposed microcompartment mechanisms are also associated with phase separation. Thus, the same or similar mechanisms are potentially associated with the insulator function of the identified DNA-binding proteins. I have included the following information on line 546: CEBPA in the identified insulator-associated DNA-binding proteins was also reported to be involved in transcriptional condensates and phase separation.

    1. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?

    Response: Snead WT et al. Molecular Cell 2019 mentioned that protein post-transcriptional modifications (PTMs) facilitate the control of molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin (Tang X et al. Nature Communications 2024). I found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Supplementary Fig. 2d). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation by PTMs. I have added the following explanation on lines 576-582: Furthermore, protein post-transcriptional modifications (PTMs) facilitate control over the molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin. We found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Fig. 4f and Supplementary Fig. 3c). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation through PTMs.

    1. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Response: Structural molecular model of cohesin-CTCF-anchored loops has been published by Li Y et al. Nature 2020. The structural conformation of CTCF and cohesin in the loops would be the cause of the directional bias of CTCF binding sites, which I mentioned in lines 531 - 535 as follows: These results suggest that the directional bias of DNA-binding sites of insulator-associated DBPs may be involved in insulator function and chromatin regulation through structural interactions among DBPs, other proteins, DNAs, and RNAs. For example, the N-terminal amino acids of CTCF have been shown to interact with RAD21 in chromatin loops. To investigate the principles underlying the architectural functions of insulator-insulator pairing interactions, two insulators, Homie and Nhomie, flanking the Drosophila even skipped locus were analyzed. Pairing interactions between the transgene Homie and the eve locus are directional. The head-to-head pairing between the transgene and endogenous Homie matches the pattern of activation (Fujioka M et al. PLoS Genetics 2016).

    Referee #3

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.

    Response: When a protein complex binds to DNA, one protein of the complex binds to the DNA directory, and the other proteins may not bind to DNA. However, the DNA motif sequence bound by the protein may be registered as the DNA-binding motif of all the proteins in the complex. The molecular structure of the complex of CTCF and Cohesin showed that both CTCF and Cohesin bind to DNA (Li Y et al. Nature 2020). I think there is a possibility that if the molecular structure of a protein complex becomes available, the previous recognition of the DNA-binding ability of a protein may be changed. Therefore, I searched the Pfam database for 99 insulator-associated DNA-binding proteins identified in this study. I found that 97 are registered as DNA-binding proteins and/or have a known DNA-binding domain, and EP300 and SIN3A do not directory bind to DNA, which was also checked by Google search. I have added the following explanation in line 249 to indicate direct and indirect DNA-binding proteins: Among 99 insulator-associated DBPs, EP300 and SIN3A do not directory interact with DNA, and thus 97 insulator-associated DBPs directory bind to DNA. I have updated the sentence in line 20 of the Abstract as follows: We discovered 97 directional and minor nondirectional motifs in human fibroblast cells that corresponded to 23 DBPs related to insulator function, CTCF, and/or other types of chromosomal transcriptional regulation reported in previous studies.

    1. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.

    Response: As the reviewer mentioned, I recognize enhancers are relatively small regions. In the paper, I intended to examine further upstream and downstream of promoter regions where enhancers are found. Therefore, I have modified the sentence in lines 917 - 919 of the Fig. 2 legend as follows: Enhancer-gene regulatory interaction regions consist of 200 bins of 10 kbp between -1 Mbp and 1 Mbp region from TSS, not including promoter.

    1. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.

    Response: Following the reviewer's advice, I have added the ChIP-seq data of H3K9me3 as a truck of the UCSC Genome Browser. The distribution of H3K9me3 signal was different from that of H3K27me3 in some regions. I also found the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions and took some screenshots of the UCSC Genome Browser of the regions around the sites in Supplementary Fig. 3b. I have modified the following sentence on lines 962 - 964 in the legend of Fig. 4: a Distribution of histone modification marks H3K27me3 (green color) and H3K9me3 (turquoise color) and transcript levels (pink color) in upstream and downstream regions of a potential insulator site (light orange color). I have also added the following result on lines 348 - 352: The same analysis was performed using H3K9me3 marks, instead of H3K27me3 (Fig. S3b). We found that the distribution of H3K9me3 signal was different from that of H3K27me3 in some regions, and discovered the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions (Fig. S3b).

    1. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    Response: The resolution of the Micro-C assay is considered to be 100 bp and above, as the human nucleome core particle contains 145 bp (and 193 bp with linker) of DNA. However, internucleosomal DNA is cleaved by endonuclease into fragments of multiples of 10 nucleotides (Pospelov VA et al. Nucleic Acids Research 1979). Highly nested focal interactions were observed (Goel VY et al. Nature Genetics 2023). Base pair resolution was reported using Micro Capture-C (Hua P et al. Nature 2021). Sub-kilobase (20 bp resolution) chromatin topology was reported using an MNase-based chromosome conformation capture (3C) approach (Aljahani A et al. Nature Communications 2022). On the other hand, Hi-C data was analyzed at 1 kb resolution. (Gu H et al. bioRxiv 2021). If the resolution of Micro-C interactions is at best at 1 kb, the binding sites of a DNA-binding protein will not show a peak around the center of the genomic locations of interaction edges. Each panel shows the number of binding sites of a specific DNA-binding protein at a specific distance from the midpoint of all chromatin interaction edges. I have modified and added the following sentences in lines 585-589: High-resolution chromatin interaction data from a Micro-C assay indicated that most of the predicted insulator-associated DBPs showed DNA-binding-site distribution peaks around chromatin interaction sites, suggesting that these DBPs are involved in chromatin interactions and that the chromatin interaction data has a high degree of resolution. Base pair resolution was reported using Micro Capture-C.

    Minor comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.

    Response: The directional bias of CTCF binding sites was identified by ChIA-pet interactions of CTCF binding sites. The analysis of the contribution scores of DNA-binding sites of proteins considering the binding sites of CTCF as an insulator showed the same tendency of directional bias of CTCF binding sites. In the analysis, to remove the false-positive prediction of DNA-binding sites, I used the binding sites that overlapped with a ChIP-seq peak of the DNA-binding protein. This result suggests that the DNA-binding sites of CTCF obtained by the current analysis have sufficient quality. Therefore, if the accuracy of prediction of DNA-binding sites is improved, althought the number of DNA-binding sites may be different, the overall tendency of the directionality of DNA-binding sites will not change and the results of this study will not change significantly.

    As for the first reference in the reviewer's comment, chromatin interaction data from Micro-C assay does not include all chromatin interactions in a cell or tissue, because it is expensive to cover all interactions. Therefore, it would be difficult to predict all chromatin interactions based on machine learning. As for the second reference in the reviewer's comment, pioneer factors such as FOXA are known to bind to closed chromatin regions, but transcription factors and DNA-binding proteins involved in chromatin interactions and insulators generally bind to open chromatin regions. The search for the DNA-binding motifs is not required in closed chromatin regions.

    1. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    Response: In the DeepLIFT paper, the authors explain that DeepLIFT is a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input (Shrikumar A et al. ICML 2017). DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

    Truly explainable AI would be able to find cause and reason, and to make choices and decisions like humans. DeepLIFT does not perform causal inferences. I did not use the term "Explainable AI" in our manuscript, but I briefly explained it in Discussion. I have added the following explanation in lines 615-620: AI (Artificial Intelligence) is considered as a black box, since the reason and cause of prediction are difficult to know. To solve this issue, tools and methods have been developed to know the reason and cause. These technologies are called Explainable AI. DeepLIFT is considered to be a tool for Explainable AI. However, DeepLIFT does not answer the reason and cause for a prediction. It calculates scores representing the contribution of the input data to the prediction.

    Furthermore, to improve the readability of the manuscript, I have included the following explanation in lines 159-165: we computed DeepLIFT scores of the input data (i.e., each binding site of the ChIP-seq data of DNA-binding proteins) in the deep leaning analysis on gene expression levels. DeepLIFT compares the importance of each input for predicting gene expression levels to its 'reference or background level' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

  10. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary:

    Osato and Hamada propose a systematic approach to identify DNA binding proteins that display directional binding. They used a modified Deep Learning method (DEcode) to investigate binding profiles of 1356 DBP from GTRD database at promoters (30 of 100bp bins around TSS) and enhancers (200 bins of 10Kb around eSNPs) and use this to predict expression of 25,071 genes in Fibroblasts, Monocytes, HMEC and NPC. This method achieves a good prediction power (Spearman correlation between predicted and actual expression of 0.74). They then use PIQ, and overlap predicted binding sites with actual ChIP-seq data to investigate the motifs of TFs that are controlling gene expression. They find 99 insulator proteins showing either a specific directional bias or minor non-directional bias, corresponding to 23 DBP previously reported to have insulator function. Of the 23 proteins they identify as regulating enhancer promoter interactions, 13 are associated with CTCF. They also show that there are significantly more insulator proteins binding sites at borders of polycomb domains, transcriptionally active or boundary regions based on chromatin interactions than other proteins.

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.
    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.
    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.
    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    Minor comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.
    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    Referee Cross-Commenting

    I would like to mention that I agree with the comments of reviewers 1 and 2.

    Significance

    General assessment:

    This is the first study to my knowledge that attempts to use Deep Learning to identify insulators and directional biases in binding. One of the limitations is that no additional methods were used to show that these DBP have directional binding bias. It is not necessarily to employ additional methods, but it would definitely strengthen the paper.

    Advancements:

    This is a useful catalogue of potential DNA binding proteins of interest, beyond just CTCF. Some known TFs are there, but also new ones are found.

    Audience:

    Basic research mainly, with particular focus on chromatin conformation and TF binding fields.

    My expertise:

    ML/AI methods in genomics, TF binding models, epigenetics and 3D chromatin interactions.

  11. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see my points below). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the following list.

    Also, I encourage the authors to integrate the current presentation of the data with other (published) data about chromatin architecture, to make more robust the claims and go deeper into the biological implications of the current work. Se my list below.

    It follows a specific list of relevant points to be addressed:

    Specific points:

    1. Introduction, line 95: CTCF appears two times, it seems redundant;
    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?
    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS;
    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details;
    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.
    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?
    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?
    8. Do the authors think that the identified DBPs could work in that way as well?
    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?
    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Significance

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, chromatin organization is an important topic in the context of a constantly expanding research field. Therefore, the work is timely and could be useful for the community. The paper appears overall well written and the figures look clear and of good quality. Nevertheless, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see list of specific points). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the above reported points.

  12. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    The study by Osato and Hamada aims at computationally identifying a set of novel putative insulator-associated DNA binding proteins (DBPs) via estimation of their contribution to the expression of genes in the same chromosome region of their binding sites (+- 1Mbp from TSS). To achieve this, the authors leverage a deep learning architecture already published via which ChIP-seq peaks of DBPs in the TSS of a given gene are used to predict its expression level in four human cell lines.

    Building on this, the authors used another tool called DeepLIFT to evaluate the weight of each DBP binding site on the final gene expression value. Hence they made the assumption that if a given DBP had an insulator function they could restrict the prediction of the gene's expression to the region included between pairs of that DBP binding sites, and evaluate the pair's motif directionality bias in the distribution of weights. They exemplify their approach's validity by the fact that they can predict the known directionality bias of CTCF/cohesin-bound sites as the highest of the lot, with the F-R orientation of the pairs the most enriched, recapitulating what already known in literature: i.e., that F-R chromatin interaction peaks are the most enriched. In addition, they find several new DBPs showing significant directionality bias; hence they could be candidates for insulation activity. They then provide correlation between these putative insulator binding sites and sites of transition between euchromatin and heterochromatin by independently using histone mark and gene expression datasets. This, of course, is not surprising because (a) there is insulation between regions with heterotypic chromatin identities, and (b) it was already known from the first papers describing insulated chromatin domains that their boundaries were well-enriched for active transcription and transcriptional regulators (e.g., Dixon et al, Nature 2012).

    Finally, they use chromatin interaction (looping) sites to check the overlap between CTCF and all other DBPs and define a subset of putative insulator DBPs not overlapping CTCF peaks, suggesting potentially new insulatory mechanisms. These factors were all known transcriptional activators, but this part of the findings carry most of the novelty in the work and have the potential of opening up new directions for research in chromatin organization.

    Overall, the methodology applied here is adequate, clear, and reproducible. The major issue, in our view, is that the entire manuscript's findings relies on the usage of deepLIFT, a tool which was not benchmarked previously or by the current study. In fact, deepLIFT is public as regards its code, and also appears as a preprint from 2017 on biorXiv and published in the Proceedings of Machine Learning Research conference. Also, this key tool was developed by the Kundaje lab (who produce high quality alogrithms), and not by the authors. Therefore, the manuscript is predominantly based on the execution of existing workflows to publicly-available data. This does not take anything away from the interesting question posed here, but at the same time does not provide the community with any new algorithm/workflow.

    Finally, although I appreciate that the authors are purely computational and have likely no capacity for experimental validation of their claims of new DBPs having insulator roles, I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning. Using this kind of data, effects on gene expression can at least be tested in regard to the authors' predictions. Moreover, in terms of validation, Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As secondary issues, we would point out that:

    • The suggested alternative transcripts function, also highlighted in the manuscript;s abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.
    • Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.
    • Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Significance

    The scientific novelty of the work lies primarily in the identification of a set of DBPs that are proposed to confer insulator activity genome-wide. This has been long sought after in human data (whilst it is well understood and defined in Drosophila). The authors produce a quantitative ranking of the putative insulation effect of these DBPs and, most importantly, go on to identify a smaller subset that are apparently non-overlapping with anchors of CTCF-cohesin loop anchors; the presence of strong motif orientation biases in many DBPs can also be of broad interest, especially those that cannot be trivially ascribable to the loop extrusion process.

    However, although these findings open the way for speculation on multiple insulation mechanisms via proteins with multiple regulatory functions, the manuscript provide no experimental or computational means to test the proposed roles of these DBPs - and, as such, this limits the potential impact of the work and mostly targets researchers in the field of genome organization that can test these findings. Having said this, if validated, this work can significantly broaden our understanding of how chromatin is organized in 3D nuclear space.

    I typically identify myself to the authors: A. Papantonis, expertise in 3D genome architecture, chromatin biology, and genomics/bioinformatics.

  13. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    I would like to thank the reviewers for their comments and interest in the manuscript and the study.

    Reviewer #1

    1) I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning.

    As the reviewer pointed out, a wet experimental validation of the results of this study would give an opportunity for more biological researchers to have an interest in the study. I plan to promote the wet experimental analysis in collaboration with biological experimental researchers as a next step of this study. The same analysis in this study can be performed in immortalized cells for CRISPR experiment (e.g. Guo Y et al. Cell 2015).

    2) Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure.

    Following the reviewer's advice, I performed the same analysis with the DNA-binding sites that do no overlap with the DNA-binding sites of CTCF and cohesin (RAD21 and SMC3) (Fig. 6 and Supplementary Fig. 4). The result showed the same tendency in the distribution of DNA-binding sites. The height of a peak on the graph became lower for some DNA-binding proteins after removing the DNA-binding sites that overlapped with those of CTCF and cohesin. I have added the following sentence on lines 427 and 817: For the insulator-associated DBPs other than CTCF, RAD21, and SMC3, the DNA-binding sites that do not overlap with those of CTCF, RND21, and SMC3 were used to examine their distribution around interaction sites.

    3) Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As suggested by the reviewer, I have added the insulator scores and boundary sites from the 4D nucleome data portal as tracks in the UCSC genome browser. The insulator scores seem to correspond to some extent to the H3K27me3 histone marks from ChIP-seq (Fig. 4a and Supplementary Fig. 3). The direction of DNA-binding sites on the genome can be shown with different colors (e.g. red and green), but the directionality is their overall tendency, and it may be difficult to notice the directionality from each binding site.

    I found that the CTCF binding sites examined by a wet experiment in the previous study may not always overlap with the boundary sites of chromatin interactions from Micro-C assay (Guo Y et al. Cell 2015). The chromatin interaction data do not include all interactions due to the high sequencing cost of the assay. The number of the boundary sites may be smaller than that of CTCF binding sites acting as insulators and/or some of the CTCF binding sites may not be locate in the boundary sites. It may be difficult for the boundary location algorithm to identify a short boundary location. Due to the limitations of the chromatin interaction data, I planned to search for insulator-associated DNA-binding proteins without using chromatin interaction data in this study. I have added the statistical summary of the analysis in lines 364-387 as follows: Overall, among 20,837 DNA-binding sites of the 97 insulator-associated proteins found at insulator sites identified by H3K27me3 histone modification marks (type 1 insulator sites), 1,315 (6%) overlapped with 264 of 17,126 5kb long boundary sites, and 6,137 (29%) overlapped with 784 of 17,126 25kb long boundary sites in HFF cells. Among 5,205 DNA-binding sites of the 97 insulator-associated DNA-binding proteins found at insulator sites identified by H3K27me3 histone modification marks and transcribed regions (type 2 insulator sites), 383 (7%) overlapped with 74 of 17,126 5-kb long boundary sites, 1,901 (37%) overlapped with 306 of 17,126 25-kb long boundary sites. Although CTCF-binding sites separate active and repressive domains, the limited number of DNA-binding sites of insulator-associated proteins found at type 1 and 2 insulator sites overlapped boundary sites identified by chromatin interaction data. Furthermore, by analyzing the regulatory regions of genes, the DNA-binding sites of the 97 insulator-associated DNA-binding proteins were found (1) at the type 1 insulator sites (based on H3K27me3 marks) in the regulatory regions of 3,170 genes, (2) at the type 2 insulator sites (based on H3K27me3 marks and gene expression levels) in the regulatory regions of 1,044 genes, and (3) at insulator sites as boundary sites identified by chromatin interaction data in the regulatory regions of 6,275 genes. The boundary sites showed the highest number of overlaps with the DNA-binding sites. Comparing the insulator sites identified by (1) and (3), 1,212 (38%) genes have both types of insulator sites. Comparing the insulator sites between (2) and (3), 389 (37%) genes have both types of insulator sites. From the comparison of insulator and boundary sites, we found that (1) or (2) types of insulator sites overlapped or were close to boundary sites identified by chromatin interaction data.

    4) The suggested alternative transcripts function, also highlighted in the manuscripts abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.

    According to the reviewer's comment, I performed the genome-wide analysis of alternative transcripts where the DNA-binding sites of insulator-associated proteins are located near splicing sites. The DNA-binding sites of insulator-associated DNA-binding proteins were found within 200 bp centered on splice sites more significantly than the other DNA-binding proteins (Fig. 4e and Table 2). I have added the following sentences on lines 397 - 404: We performed the statistical test to estimate the enrichment of insulator-associated DNA-binding sites compared to the other DNA-binding proteins, and found that the insulator-associated DNA-binding sites were significantly more abundant at splice sites than the DNA-binding sites of the other proteins (Fig 4e and Table 2; Mann‒Whitney U test, p value < 0.05). The comparison between the splice sites of both ends of first and last introns and those of other introns showed the similar statistical significance of enrichment and number of splice sites with the insulator-associated DNA-binding proteins (Table 2 and Table S9).

    5) Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.

    I believe that the Figure 1 would help researchers in other fields who are not familiar with biological phenomena and functions to understand the study. More explanation has been included in the Figures and legends of Figs. 4 and 5 to help readers outside the immediate research field understand the figures.

    6) Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Reviewer #2's comments would be related to this comment. I have introduced a more detailed explanation of the method in the Results section, as shown in the responses to Reviewer #2’s comments.

    Reviewer #2

    1) Introduction, line 95: CTCF appears two times, it seems redundant.

    On lines 91-93, I deleted the latter CTCF from the sentence "We examine the directional bias of DNA-binding sites of CTCF and insulator-associated DBPs, including those of known DBPs such as RAD21 and SMC3".

    2) Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?

    Although CTCF is known to be the main insulator protein in vertebrates, we found that 97 DNA-binding proteins including CTCF and cohesin are associated with insulator sites by modifying and developing a machine learning method to search for insulator-associated DNA-binding proteins. Most of the insulator-associated DNA-binding proteins showed the directional bias of DNA-binding motifs, suggesting that the directional bias is associated with the insulator.

    I have added the sentence in lines 96-99 as follows: Furthermore, statistical testing the contribution scores between the directional and non-directional DNA-binding sites of insulator-associated DBPs revealed that the directional sites contributed more significantly to the prediction of gene expression levels than the non-directional sites. I have revised the statement in lines 101-110 as follows: To validate these findings, we demonstrate that the DNA-binding sites of the identified insulator-associated DBPs are located within potential insulator sites, and some of the DNA-binding sites in the insulator site are found without the nearby DNA-binding sites of CTCF and cohesin. Homologous and heterologous insulator-insulator pairing interactions are orientation-dependent, as suggested by the insulator-pairing model based on experimental analysis in flies. Our method and analyses contribute to the identification of insulator- and chromatin-associated DNA-binding sites that influence EPIs and reveal novel functional roles and molecular mechanisms of DBPs associated with transcriptional condensation, phase separation and transcriptional regulation.

    3) Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS.

    On lines 121-124, to explain the procedure for the SNP of an eQTL, I have added the sentence in the Methods: "If a DNA-binding site was located within a 100-bp region around a single-nucleotide polymorphism (SNP) of an eQTL, we assumed that the DNA-binding proteins regulated the expression of the transcript corresponding to the eQTL".

    4) Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details.

    On line 119, I have included the explanation of the eQTL dataset of GTEx v8 as follows: " The eQTL data were derived from the GTEx v8 dataset, after quality control, consisting of 838 donors and 17,382 samples from 52 tissues and two cell lines”. On lines 681 and 865, I have added the filename of the eQTL data "(GTEx_Analysis_v8_eQTL.tar)".

    5) Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.

    The reviewer would mention Figure 2, not Figure 1. If so, the matrices in panels a and b in Figure 2 are equivalent. I have shown it in the figure: The same figure in panel a is rotated 90 degrees to the right. The green boxes in the matrix show the regions with the ChIP-seq peak of a DNA-binding protein overlapping with a SNP of an eQTL. I used eQTL data to associate a gene with a ChIP-seq peak that was more than 2 kb upstream and 1 kb downstream of a transcriptional start site of a gene. For each gene, the matrix was produced and the gene expression levels in cells were learned and predicted using the deep learning method. I have added the following sentences to explain the method in lines 133 - 139: Through the training, the tool learned to select the binding sites of DNA-binding proteins from ChIP-seq assays that were suitable for predicting gene expression levels in the cell types. The binding sites of a DNA-binding protein tend to be observed in common across multiple cell and tissue types. Therefore, ChIP-seq data and eQTL data in different cell and tissue types were used as input data for learning, and then the tool selected the data suitable for predicting gene expression levels in the cell types, even if the data were not obtained from the same cell types.

    6) Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?

    As suggested by the reviewer, to help readers understand the observation, I have added Supplementary Fig. S4c to show the distribution of DNA-binding sites of "CTCF, RAD21, and SMC3" and "BACH2, FOS, ATF3, NFE2, and MAFK" around chromatin interaction sites. I have modified the following sentence to indicate the figure on line 493: Although a DNA-binding-site distribution pattern around chromatin interaction sites similar to those of CTCF, RAD21, and SMC3 was observed for DBPs such as BACH2, FOS, ATF3, NFE2, and MAFK, less than 1% of the DNA-binding sites of the latter set of DBPs colocalized with CTCF, RAD21, or SMC3 in a single bin (Fig. S4c).

    In Aljahani A et al. Nature Communications 2022, we find that depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Together, our data show that loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression. Goel VY et al. Nature Genetics 2023 mentioned in the abstract: Microcompartments frequently connect enhancers and promoters and though loss of loop extrusion and inhibition of transcription disrupts some microcompartments, most are largely unaffected. These results suggested that chromatin loops can be driven by other DBPs independent of the known CTCF/Cohesin.

    I added the following sentence on lines 561-569: The depletion of cohesin causes a subtle reduction in longer-range enhancer-promoter interactions and that CTCF depletion can cause rewiring of regulatory contacts. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. Furthermore, the loop extrusion is not essential for enhancer-promoter interactions, but contributes to their robustness and specificity and to precise regulation of gene expression.

    FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates (Ji D et al. Molecular Cell 2024). CTCF have also found to form transcriptional condensate and phase separation (Lee R et al. Nucleic acids research 2022). FOS was found to be an insulator-associated DNA-binding protein in this study and is potentially involved in chromatin remodeling, transcription condensation, and phase separation with the other factors such as BACH2, ATF3, NFE2 and MAFK. I have added the following sentence on line 548: FOXA1 pioneer factor functions as an initial chromatin-binding and chromatin-remodeling factor and has been reported to form biomolecular condensates.

    7) In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?

    Goel VY et al. Nature Genetics 2023 identified highly nested and focal interactions through region capture Micro-C, which resemble fine-scale compartmental interactions and are termed microcompartments. In the section titled "Most microcompartments are robust to loss of loop extrusion," the researchers noted that a small proportion of interactions between CTCF and cohesin-bound sites exhibited significant reductions in strength when cohesin was depleted. In contrast, the majority of microcompartmental interactions remained largely unchanged under cohesin depletion. Our findings indicate that most P-P and E-P interactions, aside from a few CTCF and cohesin-bound enhancers and promoters, are likely facilitated by a compartmentalization mechanism that differs from loop extrusion. We suggest that nested, multiway, and focal microcompartments correspond to small, discrete A-compartments that arise through a compartmentalization process, potentially influenced by factors upstream of RNA Pol II initiation, such as transcription factors, co-factors, or active chromatin states. It follows that if active chromatin regions at microcompartment anchors exhibit selective "stickiness" with one another, they will tend to co-segregate, leading to the development of nested, focal interactions. This microphase separation, driven by preferential interactions among active loci within a block copolymer, may account for the striking interaction patterns we observe.

    The authors of the paper proposed several mechanisms potentially involved in microcompartments. These mechanisms may be involved in looping with insulator function. Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently (Hsieh TS et al. Nature Genetics 2022). Among the identified insulator-associated DNA-binding proteins, Maz and MyoD1 form loops without CTCF (Xiao T et al. Proc Natl Acad Sci USA 2021 ; Ortabozkoyun H et al. Nature genetics 2022 ; Wang R et al. Nature communications 2022). I have added the following sentences on lines 563-567: Another group reported that enhancer-promoter interactions and transcription are largely maintained upon depletion of CTCF, cohesin, WAPL or YY1. Instead, cohesin depletion decreased transcription factor binding to chromatin. Thus, cohesin may allow transcription factors to find and bind their targets more efficiently. I have included the following explanation on lines 574-576: Maz and MyoD1 among the identified insulator-associated DNA-binding proteins form loops without CTCF.

    As for the directionality of CTCF, if chromatin loop anchors have some structural conformation, as shown in the paper entitled "The structural basis for cohesin-CTCF-anchored loops" (Li Y et al. Nature 2020), directional DNA binding would occur similarly to CTCF binding sites. Moreover, cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops (Davidson IF et al. Nature Reviews Molecular Cell Biology 2021). Regarding loop extrusion, the ‘loop extrusion’ hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions (Guerin TM et al. EMBO Journal 2024). I have added the following sentences on lines 535-539: Cohesin complexes that interact with convergent CTCF sites, that is, the N-terminus of CTCF, might be protected from WAPL, but those that interact with divergent CTCF sites, that is, the C-terminus of CTCF, might not be protected from WAPL, which could release cohesin from chromatin and thus disrupt cohesin-mediated chromatin loops. I have included the following sentences on lines 569-574: The ‘loop extrusion’ hypothesis is motivated by in vitro observations. The experiment in yeast, in which cohesin variants that are unable to extrude DNA loops but retain the ability to topologically entrap DNA, suggested that in vivo chromatin loops are formed independently of loop extrusion. Instead, transcription promotes loop formation and acts as an extrinsic motor that extends these loops and defines their final positions.

    Another model for the regulation of gene expression by insulators is the boundary-pairing (insulator-pairing) model (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016). Molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies. Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent. I have summarized the model on lines 551-559: Other types of chromatin regulation are also expected to be related to the structural interactions of molecules. As the boundary-pairing (insulator-pairing) model, molecules bound to insulators physically pair with their partners, either head-to-head or head-to-tail, with different degrees of specificity at the termini of TADs in flies (Fig. 7). Although the experiments do not reveal how partners find each other, the mechanism unlikely requires loop extrusion. Homologous and heterologous insulator-insulator pairing interactions are central to the architectural functions of insulators. The manner of insulator-insulator interactions is orientation-dependent.

    8) Do the authors think that the identified DBPs could work in that way as well?

    The boundary-pairing (insulator-pairing) model would be applied to the insulator-associated DNA-binding proteins other than CTCF and cohesin that are involved in the loop extrusion mechanism (Bing X et al. Elife 2024) (Ke W et al. Elife 2024) (Fujioka M et al. PLoS Genetics 2016).

    Liquid-liquid phase separation was shown to occur through CTCF-mediated chromatin loops and to act as an insulator (Lee, R et al. Nucleic Acids Research 2022). Among the identified insulator-associated DNA-binding proteins, CEBPA has been found to form hubs that colocalize with transcriptional co-activators in a native cell context, which is associated with transcriptional condensate and phase separation (Christou-Kent M et al. Cell Reports 2023). The proposed microcompartment mechanisms are also associated with phase separation. Thus, the same or similar mechanisms are potentially associated with the insulator function of the identified DNA-binding proteins. I have included the following information on line 546: CEBPA in the identified insulator-associated DNA-binding proteins was also reported to be involved in transcriptional condensates and phase separation.

    9) Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?

    Snead WT et al. Molecular Cell 2019 mentioned that protein post-transcriptional modifications (PTMs) facilitate the control of molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin (Tang X et al. Nature Communications 2024). I found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Supplementary Fig. 2d). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation by PTMs. I have added the following explanation on lines 576-582: Furthermore, protein post-transcriptional modifications (PTMs) facilitate control over the molecular valency and strength of protein-protein interactions. O-GlcNAcylation as a PTM inhibits CTCF binding to chromatin. We found that the identified insulator-associated DNA-binding proteins tend to form a cluster at potential insulator sites (Fig. 4f and Supplementary Fig. 3c). These proteins may interact and actively regulate chromatin interactions, transcriptional condensation, and phase separation through PTMs.

    10) Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Structural molecular model of cohesin-CTCF-anchored loops has been published by Li Y et al. Nature 2020. The structural conformation of CTCF and cohesin in the loops would be the cause of the directional bias of CTCF binding sites, which I mentioned in lines 531 – 535 as follows: These results suggest that the directional bias of DNA-binding sites of insulator-associated DBPs may be involved in insulator function and chromatin regulation through structural interactions among DBPs, other proteins, DNAs, and RNAs. For example, the N-terminal amino acids of CTCF have been shown to interact with RAD21 in chromatin loops.

    To investigate the principles underlying the architectural functions of insulator-insulator pairing interactions, two insulators, Homie and Nhomie, flanking the Drosophila even skipped locus were analyzed. Pairing interactions between the transgene Homie and the eve locus are directional. The head-to-head pairing between the transgene and endogenous Homie matches the pattern of activation (Fujioka M et al. PLoS Genetics 2016).

    Reviewer #3

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.

    When a protein complex binds to DNA, one protein of the complex binds to the DNA directory, and the other proteins may not bind to DNA. However, the DNA motif sequence bound by the protein may be registered as the DNA-binding motif of all the proteins in the complex. The molecular structure of the complex of CTCF and Cohesin showed that both CTCF and Cohesin bind to DNA (Li Y et al. Nature 2020). I think there is a possibility that if the molecular structure of a protein complex becomes available, the previous recognition of the DNA-binding ability of a protein may be changed. Therefore, I searched the Pfam database for 99 insulator-associated DNA-binding proteins identified in this study. I found that 97 are registered as DNA-binding proteins and/or have a known DNA-binding domain, and EP300 and SIN3A do not directory bind to DNA, which was also checked by Google search. I have added the following explanation in line 249 to indicate direct and indirect DNA-binding proteins: Among 99 insulator-associated DBPs, EP300 and SIN3A do not directory interact with DNA, and thus 97 insulator-associated DBPs directory bind to DNA. I have updated the sentence in line 22 of the Abstract as follows: We discovered 97 directional and minor nondirectional motifs in human fibroblast cells that corresponded to 23 DBPs related to insulator function, CTCF, and/or other types of chromosomal transcriptional regulation reported in previous studies.

    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.

    As the reviewer mentioned, I recognize enhancers are relatively small regions. In the paper, I intended to examine further upstream and downstream of promoter regions where enhancers are found. Therefore, I have modified the sentence in lines 917 – 919 of the Fig. 2 legend as follows: Enhancer-gene regulatory interaction regions consist of 200 bins of 10 kbp between -1 Mbp and 1 Mbp region from TSS, not including promoter.

    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.

    Following the reviewer's advice, I have added the ChIP-seq data of H3K9me3 as a truck of the UCSC Genome Browser. The distribution of H3K9me3 signal was different from that of H3K27me3 in some regions. I also found the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions and took some screenshots of the UCSC Genome Browser of the regions around the sites in Supplementary Fig. 3b. I have modified the following sentence on lines 962 – 964 in the legend of Fig. 4: a Distribution of histone modification marks H3K27me3 (green color) and H3K9me3 (turquoise color) and transcript levels (pink color) in upstream and downstream regions of a potential insulator site (light orange color). I have also added the following result on lines 348 – 352: The same analysis was performed using H3K9me3 marks, instead of H3K27me3 (Fig. S3b). We found that the distribution of H3K9me3 signal was different from that of H3K27me3 in some regions, and discovered the insulator-associated DNA-binding sites close to the edges of H3K9me3 regions (Fig. S3b).

    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    The resolution of the Micro-C assay is considered to be 100 bp and above, as the human nucleome core particle contains 145 bp (and 193 bp with linker) of DNA. However, internucleosomal DNA is cleaved by endonuclease into fragments of multiples of 10 nucleotides (Pospelov VA et al. Nucleic Acids Research 1979). Highly nested focal interactions were observed (Goel VY et al. Nature Genetics 2023). Base pair resolution was reported using Micro Capture-C (Hua P et al. Nature 2021). Sub-kilobase (20 bp resolution) chromatin topology was reported using an MNase-based chromosome conformation capture (3C) approach (Aljahani A et al. Nature Communications 2022). On the other hand, Hi-C data was analyzed at 1 kb resolution. (Gu H et al. bioRxiv 2021). If the resolution of Micro-C interactions is at best at 1 kb, the binding sites of a DNA-binding protein will not show a peak around the center of the genomic locations of interaction edges. Each panel shows the number of binding sites of a specific DNA-binding protein at a specific distance from the midpoint of all chromatin interaction edges. I have modified and added the following sentences in lines 585-589: High-resolution chromatin interaction data from a Micro-C assay indicated that most of the predicted insulator-associated DBPs showed DNA-binding-site distribution peaks around chromatin interaction sites, suggesting that these DBPs are involved in chromatin interactions and that the chromatin interaction data has a high degree of resolution. Base pair resolution was reported using Micro Capture-C.

    1.PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g.,https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.

    The directional bias of CTCF binding sites was identified by ChIA-pet interactions of CTCF binding sites. The analysis of the contribution scores of DNA-binding sites of proteins considering the binding sites of CTCF as an insulator showed the same tendency of directional bias of CTCF binding sites. In the analysis, to remove the false-positive prediction of DNA-binding sites, I used the binding sites that overlapped with a ChIP-seq peak of the DNA-binding protein. This result suggests that the DNA-binding sites of CTCF obtained by the current analysis have sufficient quality. Therefore, if the accuracy of prediction of DNA-binding sites is improved, althought the number of DNA-binding sites may be different, the overall tendency of the directionality of DNA-binding sites will not change and the results of this study will not change significantly.

    As for the first reference in the reviewer's comment, chromatin interaction data from Micro-C assay does not include all chromatin interactions in a cell or tissue, because it is expensive to cover all interactions. Therefore, it would be difficult to predict all chromatin interactions based on machine learning. As for the second reference in the reviewer's comment, pioneer factors such as FOXA are known to bind to closed chromatin regions, but transcription factors and DNA-binding proteins involved in chromatin interactions and insulators generally bind to open chromatin regions. The search for the DNA-binding motifs is not required in closed chromatin regions.

    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    In the DeepLIFT paper, the authors explain that DeepLIFT is a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input (Shrikumar A et al. ICML 2017). DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

    Truly explainable AI would be able to find cause and reason, and to make choices and decisions like humans. DeepLIFT does not perform causal inferences. I did not use the term "Explainable AI" in our manuscript, but I briefly explained it in Discussion. I have added the following explanation in lines 615-620: AI (Artificial Intelligence) is considered as a black box, since the reason and cause of prediction are difficult to know. To solve this issue, tools and methods have been developed to know the reason and cause. These technologies are called Explainable AI. DeepLIFT is considered to be a tool for Explainable AI. However, DeepLIFT does not answer the reason and cause for a prediction. It calculates scores representing the contribution of the input data to the prediction.

    Furthermore, to improve the readability of the manuscript, I have included the following explanation in lines 159-165: we computed DeepLIFT scores of the input data (i.e., each binding site of the ChIP-seq data of DNA-binding proteins) in the deep leaning analysis on gene expression levels. DeepLIFT compares the importance of each input for predicting gene expression levels to its 'reference or background level' and assigns contribution scores according to the difference. DeepLIFT calculates a metric to measure the difference between an input and the reference of the input.

  14. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary:

    Osato and Hamada propose a systematic approach to identify DNA binding proteins that display directional binding. They used a modified Deep Learning method (DEcode) to investigate binding profiles of 1356 DBP from GTRD database at promoters (30 of 100bp bins around TSS) and enhancers (200 bins of 10Kb around eSNPs) and use this to predict expression of 25,071 genes in Fibroblasts, Monocytes, HMEC and NPC. This method achieves a good prediction power (Spearman correlation between predicted and actual expression of 0.74). They then use PIQ, and overlap predicted binding sites with actual ChIP-seq data to investigate the motifs of TFs that are controlling gene expression. They find 99 insulator proteins showing either a specific directional bias or minor non-directional bias, corresponding to 23 DBP previously reported to have insulator function. Of the 23 proteins they identify as regulating enhancer promoter interactions, 13 are associated with CTCF. They also show that there are significantly more insulator proteins binding sites at borders of polycomb domains, transcriptionally active or boundary regions based on chromatin interactions than other proteins.

    Major Comments:

    1. Some of these TFs do not have specific direct binding to DNA (P300, Cohesin). Since the authors are using binding motifs in their analysis workflow, I would remove those from the analysis.
    2. I am not sure if I understood correctly, by why do the authors consider enhancers spanning 2Mb (200 bins of 10Kb around eSNPs)? This seems wrong. Enhancers are relatively small regions (100bp to 1Kb) and only a very small subset form super enhancers.
    3. I think the H3K27me3 analysis was very good, but I would have liked to see also constitutive heterochromatin as well, so maybe repeat the analysis for H3K9me3.
    4. I was not sure I understood the analysis in Figure 6. The binding site is with 500bp of the interaction site, but micro-C interactions are at best at 1Kb resolution. They say they chose the centre of the interaction site, but we don't know exactly where there is the actual interaction. Also, it is not clear what they measure. Is it the number of binding sites of a specific or multiple DBP insulator proteins at a specific distance from this midpoint that they recover in all chromatin loops? Maybe I am missing something. This analysis was not very clear.

    Minor comments:

    1. PIQ does not consider TF concentration. Other methods do that and show that TF concentration improves predictions (e.g., https://www.biorxiv.org/content/10.1101/2023.07.15.549134v2 or https://pubmed.ncbi.nlm.nih.gov/37486787/). The authors should discuss how that would impact their results.
    2. DeepLIFT is a good approach to interpret complex structures of CNN, but is not truly explainable AI. I think the authors should acknowledge this.

    Referee Cross-Commenting

    I would like to mention that I agree with the comments of reviewers 1 and 2.

    Significance

    General assessment:

    This is the first study to my knowledge that attempts to use Deep Learning to identify insulators and directional biases in binding. One of the limitations is that no additional methods were used to show that these DBP have directional binding bias. It is not necessarily to employ additional methods, but it would definitely strengthen the paper.

    Advancements:

    This is a useful catalogue of potential DNA binding proteins of interest, beyond just CTCF. Some known TFs are there, but also new ones are found.

    Audience:

    Basic research mainly, with particular focus on chromatin conformation and TF binding fields.

    My expertise:

    ML/AI methods in genomics, TF binding models, epigenetics and 3D chromatin interactions.

  15. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see my points below). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the following list.

    Also, I encourage the authors to integrate the current presentation of the data with other (published) data about chromatin architecture, to make more robust the claims and go deeper into the biological implications of the current work. Se my list below.

    It follows a specific list of relevant points to be addressed:

    Specific points:

    1. Introduction, line 95: CTCF appears two times, it seems redundant;
    2. Introduction, lines 99-103: Please stress better the novelty of the work. What is the main focus? The new identified DPBs or their binding sites? What are the "novel structural and functional roles of DBPs" mentioned?
    3. Results, line 111: How do the SNPs come into the procedure? From the figures it seems the input is ChIP-seq peaks of DNBPs around the TSS;
    4. Again, are those SNPs coming from the different cell lines? Or are they from individuals w.r.t some reference genome? I suggest a general restructuring of this part to let the reader understand more easily. One option could be simplifying the details here or alternatively including all the necessary details;
    5. Figure 1: panel a and b are misleading. Is the matrix in panel a equivalent to the matrix in panel b? If not please clarify why. Maybe in b it is included the info about the SNPs? And if yes, again, what is then difference with a.
    6. Line 386-388: could the author investigate in more detail this observation? Does it mean that loops driven by other DBPs independent of the known CTCF/Cohesin? Could the author provide examples of chromatin structural data e.g. MicroC?
    7. In general, how the presented results are related to some models of chromatin architecture, e.g. loop extrusion, in which it is integrated convergent CTCF binding sites?
    8. Do the authors think that the identified DBPs could work in that way as well?
    9. Also, can the authors comment about the mechanisms those newly identified DBPs mediate contacts by active processes or equilibrium processes?
    10. Can the author provide some real examples along with published structural data (e.g. the mentioned micro-C data) to show the link between protein co-presence, directional bias and contact formation?

    Significance

    In this work, the authors describe a deep learning computational tool to identity binding motifs of DNA binding proteins associated to insulators that led to the discovery of 99 motifs related to insulation. This is in turn related to chromatin architecture and highlight the importance of directional bias in order to form chromatin loops.

    In general, chromatin organization is an important topic in the context of a constantly expanding research field. Therefore, the work is timely and could be useful for the community. The paper appears overall well written and the figures look clear and of good quality. Nevertheless, there are some aspects to be clarified and better explored to make stronger conclusions. In particular, there are some aspects to clarify in the text about the Machine Learning procedure (see list of specific points). In addition, I have some general questions about the biological implications of the discussed findings, listed in detail in the above reported points.

  16. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    The study by Osato and Hamada aims at computationally identifying a set of novel putative insulator-associated DNA binding proteins (DBPs) via estimation of their contribution to the expression of genes in the same chromosome region of their binding sites (+- 1Mbp from TSS). To achieve this, the authors leverage a deep learning architecture already published via which ChIP-seq peaks of DBPs in the TSS of a given gene are used to predict its expression level in four human cell lines.

    Building on this, the authors used another tool called DeepLIFT to evaluate the weight of each DBP binding site on the final gene expression value. Hence they made the assumption that if a given DBP had an insulator function they could restrict the prediction of the gene's expression to the region included between pairs of that DBP binding sites, and evaluate the pair's motif directionality bias in the distribution of weights. They exemplify their approach's validity by the fact that they can predict the known directionality bias of CTCF/cohesin-bound sites as the highest of the lot, with the F-R orientation of the pairs the most enriched, recapitulating what already known in literature: i.e., that F-R chromatin interaction peaks are the most enriched. In addition, they find several new DBPs showing significant directionality bias; hence they could be candidates for insulation activity. They then provide correlation between these putative insulator binding sites and sites of transition between euchromatin and heterochromatin by independently using histone mark and gene expression datasets. This, of course, is not surprising because (a) there is insulation between regions with heterotypic chromatin identities, and (b) it was already known from the first papers describing insulated chromatin domains that their boundaries were well-enriched for active transcription and transcriptional regulators (e.g., Dixon et al, Nature 2012).

    Finally, they use chromatin interaction (looping) sites to check the overlap between CTCF and all other DBPs and define a subset of putative insulator DBPs not overlapping CTCF peaks, suggesting potentially new insulatory mechanisms. These factors were all known transcriptional activators, but this part of the findings carry most of the novelty in the work and have the potential of opening up new directions for research in chromatin organization.

    Overall, the methodology applied here is adequate, clear, and reproducible. The major issue, in our view, is that the entire manuscript's findings relies on the usage of deepLIFT, a tool which was not benchmarked previously or by the current study. In fact, deepLIFT is public as regards its code, and also appears as a preprint from 2017 on biorXiv and published in the Proceedings of Machine Learning Research conference. Also, this key tool was developed by the Kundaje lab (who produce high quality alogrithms), and not by the authors. Therefore, the manuscript is predominantly based on the execution of existing workflows to publicly-available data. This does not take anything away from the interesting question posed here, but at the same time does not provide the community with any new algorithm/workflow.

    Finally, although I appreciate that the authors are purely computational and have likely no capacity for experimental validation of their claims of new DBPs having insulator roles, I would assume that there are RNA-seq and/or ChIP-seq data out there produced after knockdown of one or more of these DBPs that show directional positioning. Using this kind of data, effects on gene expression can at least be tested in regard to the authors' predictions. Moreover, in terms of validation, Figure 6 should be expanded to incorporate analysis of DBPs not overlapping CTCF/cohesin in chromatin interaction data that is important and potentially more interesting than the simple DBPs enrichment reported in the present form of the figure. Critically, I would like to see use of Micro-C/Hi-C data and ChIP-seq from these factors, where insulation scores around their directionally-bound sites show some sort of an effect like that presumed by the authors - and many such datasets are publicly-available and can be put to good use here.

    As secondary issues, we would point out that:

    • The suggested alternative transcripts function, also highlighted in the manuscript;s abstract, is only supported by visual inspection of a few cases for several putative DBPs. I believe this is insufficient to support what looks like one of the major claims of the paper when reading the abstract, and a more quantitative and genome-wide analysis must be adopted, although the authors mention it as just an 'observation'.
    • Figure 1 serves no purpose in my opinion and can be removed, while figures can generally be improved (e.g., the browser screenshots in Figs 4 and 5) for interpretability from readers outside the immediate research field.
    • Similarly, the text is rather convoluted at places and should be re-approached with more clarity for less specialized readers in mind.

    Significance

    The scientific novelty of the work lies primarily in the identification of a set of DBPs that are proposed to confer insulator activity genome-wide. This has been long sought after in human data (whilst it is well understood and defined in Drosophila). The authors produce a quantitative ranking of the putative insulation effect of these DBPs and, most importantly, go on to identify a smaller subset that are apparently non-overlapping with anchors of CTCF-cohesin loop anchors; the presence of strong motif orientation biases in many DBPs can also be of broad interest, especially those that cannot be trivially ascribable to the loop extrusion process.

    However, although these findings open the way for speculation on multiple insulation mechanisms via proteins with multiple regulatory functions, the manuscript provide no experimental or computational means to test the proposed roles of these DBPs - and, as such, this limits the potential impact of the work and mostly targets researchers in the field of genome organization that can test these findings. Having said this, if validated, this work can significantly broaden our understanding of how chromatin is organized in 3D nuclear space.

    I typically identify myself to the authors: A. Papantonis, expertise in 3D genome architecture, chromatin biology, and genomics/bioinformatics.

  17. Version published to 10.1101/2024.01.20.573595 on bioRxiv