Benchmarking tissue- and cell type-of-origin deconvolution in cell-free transcriptomics

  1. Alexis Ioannou
  2. Elias T Friman
  3. Carsten O Daub
  4. Wendy A Bickmore
  5. Simon C Biddie

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

Plasma cell-free RNA (cfRNA) reflects tissue- and cell-type-specific activity across pathological states and is a promising biomarker for organ injury and disease. Computational deconvolution methods are widely used to infer organ and cell-type contributions to cfRNA profiles. However, most were originally developed for single-tissue bulk transcriptomes and their performance in body-wide cfRNA settings, where any tissue or cell type can contribute, remains poorly characterised. Here, we present a systematic benchmarking of tissue- and cell type-of-origin deconvolution for plasma cfRNA that considers both methodological and reference-related sources of variability under realistic cfRNA simulation settings. We evaluated seven commonly used deconvolution methods across distinct algorithmic classes and multi-organ reference configurations derived from bulk and single-cell atlases. We assessed performance using simulation frameworks that model multi-organ mixtures, technical noise, and transcript degradation. We further examined deconvolution methods across multiple previously published clinical cfRNA cohorts spanning diverse disease contexts. Across both tissue- and cell-type-level analyses, deconvolution performance was strongly influenced by both method choice and reference parameters. Tissue-of-origin inference was comparatively robust across simulated and clinical datasets, recovering disease-associated organ signals and concordance with biochemical markers. In contrast, cell type-of-origin inference showed greater variability and reduced consistency across analytical settings, leading to divergent interpretations in both simulations and published clinical cfRNA cohorts. Together, these findings demonstrate that methodological and reference-related variability are major sources of uncertainty in cfRNA deconvolution, with tissue-level inference being more robust than cell-type-level inference. Our benchmarking framework provides guidance for reference selection and comparative interpretation in cfRNA deconvolution.

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

    Reviewer #1

    Evidence, reproducibility and clarity

    *Specific comments for revision - Major:

    1. The benchmarking framework relies heavily on simulated mixtures as ground truth. However, these mixtures are derived from intracellular RNA profiles and may not fully capture the biological characteristics of cfRNA, including fragmentation patterns, differential release mechanisms, and extracellular stability. This raises concerns about whether the reported performance truly reflects real-world cfRNA scenarios. The authors should explicitly discuss the limitations of this pseudo-ground truth and the potential biases introduced by the simulation design. Incorporating cfRNA-specific features or alternative validation strategies would strengthen the reliability of the conclusions.*

    Response: The reviewer highlights the framework's reliance on simulated mixtures, which may not fully capture the complexity of real-world cfRNA samples. We acknowledge this limitation and agree that simulated data cannot completely reproduce all biological and technical characteristics of cfRNA. Nevertheless, simulation-based benchmarking remains the current standard for systematically evaluating deconvolution methods because the true tissue and cell-type composition of cfRNA samples is not known. To improve the biological realism of our simulations, we incorporated two cfRNA-specific features: (i) the introduction of negative binomial noise to model technical and biological variability and (ii) the removal of rapidly degrading transcripts based on mRNA half-life, as published cfRNA data show rapidly degrading transcripts to be underrepresented.

    To further address the reviewer's comments, improving representation of simulations to cell free RNA, we will include an additional cfRNA-specific benchmarking scenario based on detectability-filtered simulations. The simulated mixtures will be restricted to genes that are consistently detected across multiple published cfRNA cohorts and diseases, thereby better reflecting the subset of transcripts that are reliably measurable in cell free RNA.

    2) The manuscript clearly demonstrates that cell type-of-origin deconvolution is substantially less robust than tissue-level inference. However, the explanation remains largely descriptive, focusing on transcriptional similarity and reference incompleteness. A deeper mechanistic analysis is needed to understand the root causes of this limitation. In particular, the authors should consider discussing the impact of collinearity between cell-type signatures, the identifiability of mixture models, and the role of signal-to-noise ratio in cfRNA data. Providing quantitative or theoretical insights would significantly enhance the contribution of the study.

    Response: The reviewer highlights that the difference in performance between tissue- and cell-type-level inference is not fully explained. Our discussion focused on the larger number of potential contributors in COO inference, its greater sensitivity in signal-to-noise, and the ability of different methods to handle correlated signatures. TOO was evaluated using approximately 30 merged tissue groups, whereas COO inference used approximately 80 merged cell-type groups. The larger COO label space increases model dimensionality and the number of potential misassignment routes. Regarding signal-to-noise ratio, noise and degradation perturbations affect both TOO and COO inference, but COO is expected to be more sensitive because the transcriptional differences separating related cell types are smaller than those separating broader tissue groups. We evaluated seven deconvolution methods with different underlying assumptions across multiple reference configurations using matched TOO and COO frameworks. The observed variation across method–reference combinations indicates that both reference design and method-specific handling of correlated signatures, model constraints, and noisy inputs influence deconvolution robustness. The reviewer also correctly points out that collinearity between cell-type signatures may contribute to the reduced performance. To investigate this, we will calculate pairwise similarities between the reference signatures for all tissue and cell-type categories in each TOO and COO reference matrix. Analysis that is being undertaken suggests higher collinearity among the COO signatures. We will further examine the relationship between signature collinearity and deconvolution error to support per-cell-type error than versus per-tissue error.

    3) The current evaluation focuses on reconstruction accuracy and correlation with biochemical markers, such as ALT. However, it remains unclear whether improved deconvolution performance translates into better clinical prediction or disease classification. Given the importance of cfRNA in biomarker discovery, the authors should consider evaluating the downstream utility of deconvolution outputs. For example, comparing predictive performance between raw cfRNA features and deconvolved proportions in classification or survival models would provide a more comprehensive assessment of practical value.

    Response: We thank the reviewer for this excellent suggestion regarding the clinical utility of deconvolution outputs. We agree that evaluating whether improved deconvolution performance translates into better disease classification, prediction, or prognostic models would be an important step toward demonstrating the practical value of cfRNA deconvolution. However, we believe that such analyses are beyond the scope of the current manuscript. The primary objective of this study was to systematically benchmark tissue- and cell-type-level cfRNA deconvolution in a whole-body setting by comparing multiple deconvolution methods across different parameter settings and reference configurations. Our focus was therefore on establishing the analytical performance and robustness of existing deconvolution approaches rather than evaluating their downstream clinical applications. Importantly, a rigorous assessment of predictive performance would require carefully curated disease-specific cohorts, appropriate clinical endpoints, and models tailored to individual clinical questions, all of which introduce additional sources of variability beyond the deconvolution task itself. We agree that comparing disease classification or survival models based on raw cfRNA expression with those incorporating deconvolved tissue or cell-type proportions is a valuable direction for future work, and we will highlight this in the Discussion.

    4) All evaluated methods belong to classical frameworks, including regression-based, Bayesian, and optimization-based approaches. Recent advances in machine learning, such as deep generative models and representation learning, are not considered in this study. The manuscript would benefit from discussing whether the observed limitations are intrinsic to the deconvolution problem or specific to current methodologies. Including a perspective on emerging approaches would improve the relevance of the work.

    Response: This point is similar to that raised by Reviewer 2 (Major Point 3), related to advanced machine learning-based deconvolution approaches absent in our benchmark. We focused on benchmarking methods that have previously been applied to cfRNA deconvolution. While deep learning methods have recently emerged for transcriptomic deconvolution in less complex settings (reviewed in https://doi.org/10.1016/j.csbj.2025.05.038), they have not yet been systematically evaluated in a body-wide cfRNA deconvolution framework. To address this point, we will expand our benchmark by including the recently developed deep learning-based deconvolution method DECODE and evaluate its performance in cell-free RNA settings. In addition, we will expand the Discussion to provide a broader perspective on emerging machine learning approaches for deconvolution, discussing whether the limitations identified in this study primarily reflect fundamental challenges of the cfRNA deconvolution problem (e.g., reference collinearity and low signal-to-noise ratio) or limitations of current methodologies, as well as the potential suitability of these emerging approaches for cell-free transcriptomic applications.

    *Specific comments for revision - Minor:

    1. The study primarily relies on mean absolute error and Pearson correlation. While these metrics are appropriate, they may not fully capture compositional differences in deconvolution results. Including additional evaluation metrics would provide a more comprehensive assessment.*

    Response: We thank the reviewer for this helpful suggestion. To provide an additional evaluation, we will include the Jensen–Shannon divergence (JSD) for each method–reference combination across the different deconvolution scenarios. As a distribution-based metric, JSD complements the existing performance measures by quantifying differences in the overall composition of the inferred tissue or cell type proportions. Because JSD depends on the number of categories in the composition, these comparisons will be performed within the same reference dataset, allowing us to assess how different deconvolution methods redistribute mass across a common set of tissues or cell types.

    2) Although the methods are described, it is not entirely clear whether default parameters were used consistently across tools or whether any tuning was performed. Providing more explicit details on parameter settings would improve reproducibility and allow fairer comparison across methods.

    Response: In the revised manuscript, we will provide more explicit details on the parameter settings to improve reproducibility. We will clarify whether the default parameters were used for each deconvolution method or whether any parameter tuning was performed, and include the relevant parameter settings in the Methods section.

    Significance

    This manuscript presents a comprehensive benchmarking study of tissue- and cell type-of-origin deconvolution methods in plasma cell-free RNA (cfRNA). The authors systematically evaluate seven widely used approaches across multiple simulated and clinical datasets, considering both methodological variability and reference-dependent effects. The inclusion of realistic simulation settings, such as noise and transcript degradation, together with validation on diverse clinical cohorts, strengthens the practical relevance of the work. The study addresses an important gap in the field, as cfRNA deconvolution is increasingly used in liquid biopsy applications but lacks standardized evaluation frameworks.

    Reviewer #2

    Evidence, reproducibility and clarity

    The authors are evaluating the performance of seven cell-type deconvolution methods using cytoplasm-free mRNA. More specifically, they are benchmarking these methods at tissue and cell type levels, assigning the tissue or cell type of origin (TOO or COO) to the mixture. Using a benchmark relevant to the cfRNA context, they demonstrate that determining TOO is simpler than estimating COO. They also demonstrate that, overall, BayesPrism is the most reliable method for deconvoluting the cfRNA signal. Finally, the study has a more translational focus, correlating cfRNA deconvolution from a published dataset with biomarkers linked to tissue damage. The authors found that the results of RNA deconvolution are linked to biomarkers and could potentially be used to retrieve the disease-associated signal produced by injured tissue. COO is less correlated with the biomarkers than TOO, and BayesPrism outperformed the other deconvolution tools tested, strengthening the previous benchmarks.

    *Major comments: The manuscript is well structured and written relatively clearly. My main concern about the study is the choices made in its design. It is not always clear from the manuscript or the figures why these choices were made. While these choices are correct, their justification is either absent or poorly stated. This includes:

    • the removal of 10-40% of rapidly degrading rRNA from the signature matrices
    • central vs random (5, 10) reference profiles
    • maximum signature sizes
    • inner/outer merges on figure S7.*

    Response: In the revised version, we will provide rationale for the design decisions used in the manuscript. We will provide clearer justification for the removal of rapidly degrading mRNA from the signature matrices, the use of central versus random reference profiles, the maximum signature sizes, and the inner versus outer merge strategies used in Figure S7.

    Another concern is that most end users are more interested in the relative differential abundance of cell types/tissues between samples than in the absolute proportions of cell types/tissues. Could the author create a figure showing which method can identify the cell types/tissues that are differentially present between samples, as the results can differ from the overall accuracy?

    Response: We thank the reviewer for this valuable suggestion. We agree that, in many applications, users are more interested in relative differences in tissue or cell type abundance between samples. In the current manuscript, relative abundances of tissues and cell types of interest are presented in several figures, including Figures 6 and 7 and Supplementary Figures S12, S14, and S15. To address the reviewer's suggestion more directly, we will include an additional figure in the revised manuscript showing the distribution of estimated proportions for each tissue or cell type across individual samples, stratified by disease group and by study cohort. This will facilitate the identification of tissues and cell types that are differentially abundant between samples and complement the overall benchmarking results.

    The panel of chosen deconvolution methods is fine. However, I would add Scaden or preferably DECODE to complete the methods with a deep learning approach. The analyses seem reproducible, and the code is already available and well organized.

    Response: This point is similar to that raised by Reviewer 1 (Major Point 4). We will include DECODE in the revised benchmark and are currently training and evaluating DECODE for multi-organ cell-free RNA deconvolution task. Compared with Scaden, DECODE is a more suitable deep learning approach for this application. It is computationally efficient, which is advantageous for this more complex task, and, by design, can accommodate heterogeneous reference datasets with substantial batch effects.

    Minor comments:

    There are too many commas in the affiliations.

    Response: This will be corrected in the resubmission.

    The manuscript needs to cite Svenningsen et al. (2024): https://doi.org/10.1002/jev2.12511, as, to my knowledge, it is the only previous study of deconvolution from extracellular RNA.

    Response: We will cite this reference in the revised manuscript.

    Correlation figures such as S2A should be colored by cell type/tissue. If this results in too many colours, some cell types should be merged.

    Response: This will be corrected in the resubmission. We will highlighting selected tissues and cell type examples. Including all 30 tissues and 80 cell types would make the figure uninterpretable, and merging would undermine the individual tissue and cell-type interpretation.

    Figure 4B: It is difficult to assess what constitutes a good result. Please add points of the ground truth if relevant.

    Response: Figure 4B shows the estimated proportions of brain cell types following deconvolution of bulk RNA-sequencing data derived from brain tissue. The expected total proportion across all inferred brain cell types is 100%. The purpose of this analysis is to evaluate how different method–reference combinations using the brain-augmented reference influence the inferred composition of brain cell types. We will revise the figure legend to clarify this.

    There is a repetition in the Fig S7 legend (augmented with augmented with).

    Response: This will be corrected in the resubmission.

    Significance

    *The use of deconvolution on cfRNA is novel and could contribute to bridging the gap between the development of deconvolution methods and their application, for example in a clinical context. This demonstrates that cell type (or tissue) deconvolution could be employed in personalized medicine applications.

    The study also provides insights into how to benchmark deconvolution in the context of cfRNA, such as depleting low-half-life mRNA.

    While this work does not present any new deconvolution methods, datasets or benchmarks outside the context of cfRNA, I believe its contribution is significant enough to be published and to reach a wide audience, ranging from deconvolution method developers to bioinformaticians working in the field of personalized medicine.

    An interesting development of this work would be to further close the gap between benchmarks and the clinical use of cfRNA deconvolution by providing clearer usage guidance and testing it experimentally. Expertise of the reviewer: OMICS analyses, cell-type deconvolution*

  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 #2

    Evidence, reproducibility and clarity

    The authors are evaluating the performance of seven cell-type deconvolution methods using cytoplasm-free mRNA. More specifically, they are benchmarking these methods at tissue and cell type levels, assigning the tissue or cell type of origin (TOO or COO) to the mixture. Using a benchmark relevant to the cfRNA context, they demonstrate that determining TOO is simpler than estimating COO. They also demonstrate that, overall, BayesPrism is the most reliable method for deconvoluting the cfRNA signal. Finally, the study has a more translational focus, correlating cfRNA deconvolution from a published dataset with biomarkers linked to tissue damage. The authors found that the results of RNA deconvolution are linked to biomarkers and could potentially be used to retrieve the disease-associated signal produced by injured tissue. COO is less correlated with the biomarkers than TOO, and BayesPrism outperformed the other deconvolution tools tested, strengthening the previous benchmarks.

    Major comments:

    The manuscript is well structured and written relatively clearly. My main concern about the study is the choices made in its design. It is not always clear from the manuscript or the figures why these choices were made. While these choices are correct, their justification is either absent or poorly stated. This includes:

    • the removal of 10-40% of rapidly degrading rRNA from the signature matrices
    • central vs random (5, 10) reference profiles
    • maximum signature sizes
    • inner/outer merges on figure S7.

    Another concern is that most end users are more interested in the relative differential abundance of cell types/tissues between samples than in the absolute proportions of cell types/tissues. Could the author create a figure showing which method can identify the cell types/tissues that are differentially present between samples, as the results can differ from the overall accuracy?

    The panel of chosen deconvolution methods is fine. However, I would add Scaden or preferably DECODE to complete the methods with a deep learning approach. The analyses seem reproducible, and the code is already available and well organized.

    Minor comments:

    There are too many commas in the affiliations.

    The manuscript needs to cite Svenningsen et al. (2024): https://doi.org/10.1002/jev2.12511, as, to my knowledge, it is the only previous study of deconvolution from extracellular RNA.

    Correlation figures such as S2A should be colored by cell type/tissue. If this results in too many colours, some cell types should be merged.

    Figure 4B: It is difficult to assess what constitutes a good result. Please add points of the ground truth if relevant.

    There is a repetition in the Fig S7 legend (augmented with augmented with).

    Referees cross-commenting

    I mostly agree with Reviewer #1. I would not consider the following two points to be mandatory:

    • A deeper mechanistic analysis to understand the root causes of the difference between COO and TOO. While the question of why is of the utmost interest, I feel it is outside the scope of the manuscript. However, the authors could discuss the potential causes of these differences in more detail in the discussion and leave the question open for future work in this domain.
    • The default parameters of the tools used are clear in the GitHub repository associated with the manuscript. I will leave it to the editor to decide whether it is sufficient or whether the methods section should be clearer, as Reviewer #1 suggests. I have no strong opinion about it.

    Significance

    The use of deconvolution on cfRNA is novel and could contribute to bridging the gap between the development of deconvolution methods and their application, for example in a clinical context. This demonstrates that cell type (or tissue) deconvolution could be employed in personalized medicine applications.

    The study also provides insights into how to benchmark deconvolution in the context of cfRNA, such as depleting low-half-life mRNA.

    While this work does not present any new deconvolution methods, datasets or benchmarks outside the context of cfRNA, I believe its contribution is significant enough to be published and to reach a wide audience, ranging from deconvolution method developers to bioinformaticians working in the field of personalized medicine.

    An interesting development of this work would be to further close the gap between benchmarks and the clinical use of cfRNA deconvolution by providing clearer usage guidance and testing it experimentally.

    Expertise of the reviewer: OMICS analyses, cell-type deconvolution

  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 #1

    Evidence, reproducibility and clarity

    Specific comments for revision - Major:

    1. The benchmarking framework relies heavily on simulated mixtures as ground truth. However, these mixtures are derived from intracellular RNA profiles and may not fully capture the biological characteristics of cfRNA, including fragmentation patterns, differential release mechanisms, and extracellular stability. This raises concerns about whether the reported performance truly reflects real-world cfRNA scenarios. The authors should explicitly discuss the limitations of this pseudo-ground truth and the potential biases introduced by the simulation design. Incorporating cfRNA-specific features or alternative validation strategies would strengthen the reliability of the conclusions.

    2. The manuscript clearly demonstrates that cell type-of-origin deconvolution is substantially less robust than tissue-level inference. However, the explanation remains largely descriptive, focusing on transcriptional similarity and reference incompleteness. A deeper mechanistic analysis is needed to understand the root causes of this limitation. In particular, the authors should consider discussing the impact of collinearity between cell-type signatures, the identifiability of mixture models, and the role of signal-to-noise ratio in cfRNA data. Providing quantitative or theoretical insights would significantly enhance the contribution of the study.

    3. The current evaluation focuses on reconstruction accuracy and correlation with biochemical markers, such as ALT. However, it remains unclear whether improved deconvolution performance translates into better clinical prediction or disease classification. Given the importance of cfRNA in biomarker discovery, the authors should consider evaluating the downstream utility of deconvolution outputs. For example, comparing predictive performance between raw cfRNA features and deconvolved proportions in classification or survival models would provide a more comprehensive assessment of practical value.

    4. All evaluated methods belong to classical frameworks, including regression-based, Bayesian, and optimization-based approaches. Recent advances in machine learning, such as deep generative models and representation learning, are not considered in this study. The manuscript would benefit from discussing whether the observed limitations are intrinsic to the deconvolution problem or specific to current methodologies. Including a perspective on emerging approaches would improve the relevance of the work.

    Specific comments for revision - Minor:

    1. The study primarily relies on mean absolute error and Pearson correlation. While these metrics are appropriate, they may not fully capture compositional differences in deconvolution results. Including additional evaluation metrics would provide a more comprehensive assessment.

    2. Although the methods are described, it is not entirely clear whether default parameters were used consistently across tools or whether any tuning was performed. Providing more explicit details on parameter settings would improve reproducibility and allow fairer comparison across methods.

    Significance

    This manuscript presents a comprehensive benchmarking study of tissue- and cell type-of-origin deconvolution methods in plasma cell-free RNA (cfRNA). The authors systematically evaluate seven widely used approaches across multiple simulated and clinical datasets, considering both methodological variability and reference-dependent effects. The inclusion of realistic simulation settings, such as noise and transcript degradation, together with validation on diverse clinical cohorts, strengthens the practical relevance of the work. The study addresses an important gap in the field, as cfRNA deconvolution is increasingly used in liquid biopsy applications but lacks standardized evaluation frameworks.

  4. Version published to 10.64898/2026.03.05.709833 on bioRxiv