TSvelo: Comprehensive RNA velocity by modeling the cascade of gene regulation, transcription and splicing

  1. Jiachen Li
  2. Zhe Wang
  3. Hong-Bin Shen
  4. Ye Yuan

  • Curated by eLife

    eLife logo

    eLife Assessment

    This study presents a valuable tool named TSvelo, a computational framework for RNA velocity inference that models transcriptional regulation and gene-specific splicing. The evidence supporting the claims of the authors is solid, although elaboration of the computational benchmark and datasets would have strengthened the study. The work will be of interest to computational scientists working in the field of RNA biology.

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

RNA velocity approaches fit gene dynamics and infer cell fate by modeling the splicing process using single-cell RNA sequencing (scRNA-seq) data. However, due to short time scale of splicing, high noise and large complexity of data, existing RNA velocity methods often fail to precisely capture the complex velocity dynamics for individual gene and single cell, which makes its downstream analysis less reliable and less robust. We propose TSvelo , a comprehensive RNA velo city mathematics framework that can model the cascade of gene regulation, T ranscription and S plicing using highly interpretable neural Ordinary Differential Equations (ODEs). TSvelo can precisely capture the transcription-unspliced-spliced 3D dynamics of all genes simultaneously, infer unified latent time shared by genes within single cell, detect key gene regulatory relations and be applied to multi-lineage datasets. Experiments on six scRNA-seq datasets, including two multi-lineage datasets, demonstrate TSvelo’s superiority.

Article activity feed

  1. eLife

    eLife Assessment

    This study presents a valuable tool named TSvelo, a computational framework for RNA velocity inference that models transcriptional regulation and gene-specific splicing. The evidence supporting the claims of the authors is solid, although elaboration of the computational benchmark and datasets would have strengthened the study. The work will be of interest to computational scientists working in the field of RNA biology.

  2. eLife

    Reviewer #1 (Public review):

    Summary:

    In the paper, the authors propose a new RNA velocity method, TSvelo, which predicts the transcription rate linearly based on the expression of RNA levels of transcription factors. This framework is an extension of its recent work TFvelo by including unspliced reads and designing a coherent neuralODE framework. Improved performance was demonstrated in six diverse datasets.

    Strengths:

    Overall, this method introduces innovative solutions to link cell differentiation and gene regulation, with a balance between model complexity (neuralODE) and interpretability (raw gene space).

    Weaknesses:

    While it seems to provide convincing results, there are multiple technical concerns for the authors to clarify and double-check.

    (1) The authors should clarify and discuss the TF-target map: here, the TF-target genes map is predefined by the TF binding's ChIP-seq data. This annotation is largely incomplete and mostly compiled from a set of bulk tissues. Therefore, for a certain population, the TF-target relation may change. This requires clarification and discussion, possibly exploring how to address this in the model. In addition, a regulon database could be added, e.g., DoRothEA?

    (2) The authors should clarify how example genes are selected. This is particularly unclear in Figure 2d.

    (3) The authors should clarify confidence in the statement in lines 179-180, that ANXA4 should initially decrease. This is particularly concerning, as TSvelo didn't capture the cell cycle transitions well during the initial part.

    (4) A support reference should be added for the statement in line 260 that "neuron migrations are inside-out manner". There is no reference supporting this, and this statement is critical for the model assessment.

    (5) The comparison to scMultiomics data is particularly interesting, as MultiVelo uses ATAC data to predict the transcription rate. It would be very insightful to add a direct comparison of the estimated transcription rate between using ATAC and directly using TFs' RNA expressions.

    (6) In Figure 6g, it should be clarified how the lineage was determined. Did the authors use the LARRY barcodes, predicted cell fate, or any other methods? Here, the best way is probably using the LARRY barcodes for individual clones.

  3. eLife

    Reviewer #2 (Public review):

    Summary:

    Li et al. propose TSvelo, a computational framework for RNA velocity inference that models transcriptional regulation and gene-specific splicing using a neural ODE approach. The method is intended to improve trajectory reconstruction and capture dynamic gene expression changes in scRNA-seq data. However, the manuscript in its current form falls short in several critical areas, including rigorous validation, quantitative benchmarking, clarity of definitions, proper use of prior knowledge, and interpretive caution. Many of the authors' claims are not fully supported by the evidence.

    Major comments:

    (1) Modeling comments

    (a) Lines 512-513: How does the U-to-S delay validate the accuracy of pseudotime? Using only a single gene as an example is not sufficient for "validation."

    (b) Lines 512-518: The authors propose a strategy for selecting the initial state, but do not benchmark how accurate this selection procedure is, nor do they provide sufficient rationale. While some genes may indeed exhibit U-to-S delay during lineage differentiation, why does the highest U-to-S delay score indicate the correct initiation states? Please provide mathematical justification and demonstrate accuracy beyond using a single gene example. Maybe a simulation with ground truth could help here, too.

    (c) Equation (8): The formulation looks to be incorrect. If $$W \in \mathbb{R}^{G\times G}$$ and $$W' - \Gamma' \in \mathbb{R}^{K\times K}$$, how can they be aligned within the same row? Please clarify.

    (d) The use of prior knowledge graphs from ENCODE or ChEA to constrain regulation raises concerns. Much of the regulatory information in these databases comes from cell lines. How can such cell-line-based regulation be reliably applied to primary tissues, as is done throughout the manuscript? Additional experiments are needed to test the robustness of TSvelo with respect to prior knowledge.

    (e) Lines 579-580: How is the grid search performed? More methodological details are required. If an existing method was used, please provide a citation.

    (2) Application on pancreatic endocrine datasets

    (a) Lines 140-141: What is the definition of the final pseudotime-fitted time t or velocity pseudotime?

    (b) Lines 143-144: The use of the velocity consistency metric to benchmark methods in multi-lineage datasets is incorrect. In multi-lineage differentiation systems, cells (e.g., those in fate priming stages) may inherently show inconsistency in their velocity. Thus, it is difficult to distinguish inconsistency caused by estimation error from that arising from biological signals. Velocity consistency metrics are only appropriate in systems with unidirectional trajectories (e.g., cell cycling). The abnormally high consistency values here raise concerns about whether the estimated velocities meaningfully capture lineage differences.

    (c) The improvement of TSvelo over other methods in terms of cross-boundary direction correctness looks marginal; a statistical test would help to assess its significance.

    (d) Lines 177-178: Based on the figure, TSvelo does not appear to clearly distinguish cell types. A quantitative metric, such as Adjusted Rand Index (ARI), should be provided.

    (e) Lines 179-183: The claim that traditional methods cannot capture dynamics in the unspliced-spliced phase portrait is vague. What specific aspect is not captured-the fitted values or something else? Evidence is lacking. Please provide a detailed explanation and quantitative metrics to support this claim.

    (3) Application to gastrulation erythroid datasets

    (a) Lines 191-194: The observation that velocity genes are enriched for erythropoiesis-related pathways is trivial, since the analysis is restricted to highly variable genes (HVGs) from an erythropoiesis dataset. This enrichment is expected and therefore not informative.

    (b) Lines 227-228: It remains unclear how TSvelo "accurately captures the dynamics." What is the definition of dynamics in this context? Figure 3g shows unspliced/spliced vs. fitted time plots and phase portraits, but without a quantitative definition or measure, the claim of superiority cannot be supported. Visualization of a single gene is insufficient; a systematic and quantitative analysis is needed.

    (4) Application to the mouse brain and other datasets

    (a) Lines 280-281: The authors cannot claim that velocity streams are smoother in TSvelo than in Multivelo based solely on 2D visualization. Similarly, claiming that one model predicts the correct differentiation trajectory from a 2D projection is over-interpretation, as has been discussed in prior literature see PMID: 37885016.

    (b) Lines 304-306: Beyond transcriptional signal estimation, how is regulation inferred solely from scRNA-seq data validated, especially compared with scATAC-seq data? Are there cases where transcriptome-based regulatory inference is supported by epigenomic evidence, thereby demonstrating TSvelo's GRN inference accuracy?

    (c) The claim that TSvelo can model multi-lineage datasets hinges on its use of PAGA for lineage segmentation, followed by independent modeling of dynamics within each subset. However, the procedure for merging results across subsets remains unclear.

  4. eLife

    Reviewer #3 (Public review):

    Despite the abundance of RNA velocity tools, there are still major limitations, and there is strong skepticism about the results these methods lead to. In this paper, the authors try to address some limitations of current RNA velocity approaches by proposing a unified framework to jointly infer transcriptional and splicing dynamics. The method is then benchmarked on 6 real datasets against the most popular RNA velocity tools.

    While the approach has the potential to be of interest for the field, and may present improvements compared to existing approaches, there are some major limitations that should be addressed, particularly concerning the benchmark (see major comment 1).

    Major comments:

    (1) My main criticism concerns the benchmarking: real data lack a ground truth, and are absolutely not ideal for comparing methods, because one can only speculate what results appear to be more plausible.
    A solid and extensive simulation study, which covers various scenarios and possibly distinct data-generating models, is needed for comparing approaches. The authors should check, for example, the simulation studies in the BayVel approach (Section 4, BayVel: A Bayesian Framework for RNA Velocity Estimation in Single-Cell Transcriptomics). Clearly, all methods should be included in the simulation.

    (2) Related to the above: since a ground truth is missing, the real data analyses need to be interpreted with caution. I recommend avoiding strong statements, such as "successfully captures the correct gene dynamics", or "accurately infer", in favour of milder statements supported by the data, such as "... aligns with the biological processes described" (as in page 12), or "results are compatible with current biological knowledge", etc...

    (3) Many methods perform RNA velocity analyses. While there is a brief description, I think it'd be useful to have a schematic summary (e.g., via a Table) of the main conceptual, mathematical, and computational characteristics of each approach.

    (4) Related to the above: I struggled to identify the main conceptual novelty of TSvelo, compared to existing approaches. I recommend explaining this aspect more extensively.

    (5) A computational benchmark is missing; I'd appreciate seeing the runtime and memory cost of all methods in a couple of datasets.

    (6) I think BayVel (mentioned above) should be added to the list of competing methods (both in the text and in the benchmarks). The package can be found here: https://github.com/elenasabbioni/BayVel_pkgJulia .

  5. eLife

    Author response:

    Reviewer #1:

    We appreciate the reviewer’s positive assessment of TSvelo and their helpful technical comments. In the revised manuscript, we will:

    (1) Provide a clearer discussion of TF–target annotations, their limitations, and potential integration of additional databases.

    (2) Clarify the rationale for example-gene selection (e.g., in Fig. 2d).

    (3) Re-evaluate and temper the interpretation regarding ANXA4 and early-stage cell-cycle transitions.

    (4) Add appropriate references supporting neuronal inside-out migration.

    (5) Include additional analysis comparing TF-based transcription rate estimation with ATAC-based estimates from MultiVelo.

    (6) Clarify how lineages were determined in Fig. 6g and incorporate barcode-based validation where applicable.

    (7) Correct all typographical errors noted.

    Reviewer #2:

    We appreciate the reviewer’s careful examination of modeling, benchmarking, and interpretation. To address these concerns, we will:

    (1) Expand the methodological justification for initial-state selection, add simulations with ground truth, and evaluate U-to-S delay more broadly across genes.

    (2) Clarify matrix formulations and ensure consistency in notation (e.g., Eq. 8).

    (3) Assess robustness to prior-knowledge graphs and evaluate alternatives beyond ENCODE/ChEA.

    (4) Add methodological details on parameter search.

    (5) Improve benchmarking on pancreatic endocrine datasets by including clear definitions of velocity pseudotime, ARI for cell-type separation, quantitative evaluation of phase-portrait fits, and appropriate interpretation of consistency metrics for multi-lineage systems.

    (6) Reframe claims about “accurate” or “correct” predictions where evidence is qualitative and strengthen quantitative support where possible.

    (8) Clarify lineage segmentation and merging when applying PAGA-guided multi-lineage modeling.

    Reviewer #3:

    We thank the reviewer for highlighting the need for more rigorous benchmarking and conceptual clarity. In response, we will:

    (1) Conduct an expanded simulation study incorporating different data-generating models.

    (2) Revise all strong claims to more cautious, evidence-based language.

    (3) Add a concise table summarizing conceptual and computational differences among RNA-velocity frameworks.

    (4) More clearly articulate the conceptual novelty of TSvelo relative to existing approaches.

    (5) Include runtime and memory benchmarks across representative datasets.

    (6) Explore additional methods in conceptual comparisons and benchmarking analyses.We appreciate the reviewers’ thoughtful input and agree that the suggested analyses and clarifications will significantly improve the rigor and clarity of the manuscript. We will incorporate all recommended revisions in the resubmission and provide a full, detailed, point-by-point response at that time.

  6. Version published to 10.1101/2024.12.24.630058 on bioRxiv