pith. sign in

arxiv: 2606.00685 · v1 · pith:T7FOKI25new · submitted 2026-05-30 · 💻 cs.LG

Prior-Guided Multi-Omic Transformers for Single-Cell Gene Regulatory Network Inference

Tianyang Xu , Tianci Liu , Niraj Rayamajhi , Ryan Patrick , Kranthi Varala , Ying Li , Jing Gao This is my paper

Pith reviewed 2026-06-28 19:23 UTC · model grok-4.3

classification 💻 cs.LG
keywords gene regulatory networkssingle-cell multi-omicstransformer modelsprior-guided learningGRN inferencechromatin accessibilityweak supervisioncross-attention
0
0 comments X

The pith

A Transformer learns data-driven gene-peak links from single-cell data and uses bulk-derived priors as noisy supervision to reconstruct more accurate gene regulatory networks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces EpiAwareNet to reconstruct gene regulatory networks from paired single-cell transcriptomic and chromatin accessibility measurements. In the first stage a cross-attention module builds joint gene-peak representations that aggregate accessibility signals in a gene-specific way rather than relying on fixed peak-to-gene assignments. In the second stage a bulk-derived GRN prior is added as noisy positive edges to supply weak supervision when labeled interactions are scarce. Experiments show the resulting networks recover more known regulatory interactions than single- or multi-omic baselines. A reader would care because better GRN maps clarify how transcription factors control cell states and disease processes.

Core claim

EpiAwareNet reconstructs GRNs from paired single-cell transcriptomic and chromatin accessibility data by first learning joint gene-peak representations through a gene-peak cross-attention module that performs data-driven aggregation of accessibility signals, then incorporating a bulk-derived GRN prior as noisy positive edges to provide weak supervision under label scarcity, yielding improved recovery of known regulatory interactions compared with representative baselines.

What carries the argument

Gene-peak cross-attention module for adaptive joint representations, combined with incorporation of bulk GRN prior as noisy positive edges for weak supervision.

If this is right

  • EpiAwareNet improves GRN reconstruction performance over representative single- and multi-omic baselines.
  • The inferred GRNs exhibit greater biological plausibility through improved recovery of known regulatory interactions.
  • Lightweight biological priors from bulk data can guide single-cell GRN inference when paired with adaptive cross-modal representation learning.
  • The two-stage design remains robust to noise in the supplied prior while operating under label scarcity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could reduce reliance on large labeled single-cell datasets by leveraging existing bulk resources for other cell types or conditions.
  • Similar prior-guided cross-attention designs might apply to additional multi-omic prediction tasks such as cell-type-specific chromatin modeling.
  • If the noise-robustness property holds, the framework could be tested on priors derived from different tissues or species to measure transferability.

Load-bearing premise

The bulk-derived GRN prior, treated as noisy positive edges, supplies useful weak supervision that improves performance without introducing systematic bias.

What would settle it

If adding the bulk prior to the model consistently lowers recovery of held-out known interactions or increases false positives relative to the single-cell-only version, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.00685 by Jing Gao, Kranthi Varala, Niraj Rayamajhi, Ryan Patrick, Tianci Liu, Tianyang Xu, Ying Li.

Figure 1
Figure 1. Figure 1: Overview of the EpiAwareNet framework. Stage 1 encodes paired scRNA-seq/scATAC-seq with a prior-guided multi [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: PR, ROC, Precision@K, and Hit@K curves for pN and mN datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Robustness to noise in the bulk-derived GRN prior. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case study on tomato pN: For unseen regulators [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: PR, ROC, Precision@K, and Hit@K curves for PBMC and Mouse Brain datasets. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scalability of EpiAwareNet on tomato pN. On tomato pN, end-to-end runtime increases smoothly and approx￾imately linearly as the number of cells 𝑁 or genes 𝐺 increases under fixed settings ( [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
read the original abstract

Gene regulatory networks (GRNs) capture transcription factor-target interactions and are central to understanding cell-state regulation and disease. Reconstructing GRNs from paired single-cell transcriptomic and chromatin accessibility data is promising but challenging: scATAC is extremely sparse, and most methods rely on fixed peak-to-gene links and weak supervision. We present EpiAwareNet, a prior-guided multi-omic Transformer framework that reconstructs GRNs from paired single-cell data using only lightweight biological priors. In Stage 1, EpiAwareNet learns joint gene-peak representations with a gene-peak cross-attention module, enabling data-driven, gene-specific aggregation of accessibility signals rather than hard-coded peak-to-gene assignments. In Stage 2, EpiAwareNet incorporates a bulk-derived GRN prior as noisy positive edges to provide weak supervision under label scarcity, refining regulatory scores while remaining robust to prior noise. In our experiments, EpiAwareNet improves GRN reconstruction over representative single- and multi-omic baselines and yields GRNs with greater biological plausibility, such as improved recovery of known regulatory interactions, suggesting that lightweight biological priors from bulk data can effectively guide single-cell GRN inference when combined with adaptive cross-modal representation learning. Code and data will be available at https://github.com/tianyang-x/EpiAwareNet_pub.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents EpiAwareNet, a two-stage prior-guided multi-omic Transformer for inferring gene regulatory networks (GRNs) from paired single-cell RNA-seq and ATAC-seq data. Stage 1 employs a gene-peak cross-attention module to learn joint representations and perform data-driven aggregation of accessibility signals, avoiding fixed peak-to-gene links. Stage 2 incorporates a bulk-derived GRN prior as noisy positive edges to supply weak supervision under label scarcity, refining regulatory scores while claiming robustness to prior noise. Experiments report improved GRN reconstruction over single- and multi-omic baselines together with greater biological plausibility via better recovery of known regulatory interactions.

Significance. If the reported gains are shown to arise from the paired single-cell data rather than bulk prior overlap, the framework would offer a practical route to leverage lightweight external priors for single-cell GRN tasks. The adaptive cross-attention mechanism addresses a known limitation of hard-coded peak-to-gene mappings, and the emphasis on robustness to noisy priors is a useful design principle.

major comments (2)
  1. [Stage 2] Stage 2 (prior incorporation): the claim that the bulk GRN prior supplies useful weak supervision without introducing systematic bias toward non-cell-type-specific interactions is load-bearing for the plausibility results, yet the abstract and method description provide no quantitative ablation that isolates edges supported by the paired scRNA/scATAC data alone versus those recovered primarily through prior overlap.
  2. [Experiments] Experimental evaluation: the reported improvements in GRN reconstruction and recovery of known interactions lack visible details on data splits, multiple-testing correction, or statistical tests that would confirm the gains survive correction; without these, it is unclear whether the central claim of outperformance is robust.
minor comments (1)
  1. The abstract states that code and data will be available at a GitHub link; confirming that the repository contains the exact data splits and evaluation scripts used in the reported experiments would strengthen reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below, clarifying our approach and outlining planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Stage 2] Stage 2 (prior incorporation): the claim that the bulk GRN prior supplies useful weak supervision without introducing systematic bias toward non-cell-type-specific interactions is load-bearing for the plausibility results, yet the abstract and method description provide no quantitative ablation that isolates edges supported by the paired scRNA/scATAC data alone versus those recovered primarily through prior overlap.

    Authors: We agree that an explicit ablation isolating the contribution of the paired single-cell data from the bulk prior is important to substantiate the claim of robustness and to rule out systematic bias from prior overlap. In the revised manuscript we will add a new ablation experiment comparing (i) the full EpiAwareNet model, (ii) a variant trained without the prior (using only the cross-attention representations from paired scRNA/scATAC), and (iii) a prior-only baseline. We will report the fraction of recovered edges unique to the single-cell data, the overlap with the bulk prior, and performance on cell-type-specific benchmarks to quantify the added value of the paired data. revision: yes

  2. Referee: [Experiments] Experimental evaluation: the reported improvements in GRN reconstruction and recovery of known interactions lack visible details on data splits, multiple-testing correction, or statistical tests that would confirm the gains survive correction; without these, it is unclear whether the central claim of outperformance is robust.

    Authors: We acknowledge that the current manuscript does not provide sufficient detail on the evaluation protocol. In the revision we will expand the Experiments section to include: (1) explicit description of data splits (e.g., cell-wise or gene-wise train/validation/test partitions and how they avoid leakage), (2) the multiple-testing correction procedure applied to reported metrics, and (3) the statistical tests (with p-values) used to compare EpiAwareNet against baselines. These additions will allow readers to assess the robustness of the reported gains. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML framework with external validation

full rationale

The paper describes a two-stage Transformer model for GRN inference that learns representations via cross-attention in Stage 1 and applies a bulk-derived prior as weak supervision in Stage 2. No equations, derivations, or fitted parameters are presented that reduce any reported improvement or plausibility metric to a quantity defined by the prior itself. Claims rest on comparisons to external baselines and recovery of known interactions, with no self-citation load-bearing steps or self-definitional constructions visible in the provided text. The approach is self-contained against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the method implicitly relies on standard Transformer training assumptions and the domain assumption that bulk priors remain informative despite noise.

pith-pipeline@v0.9.1-grok · 5784 in / 1246 out tokens · 18032 ms · 2026-06-28T19:23:51.337460+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [1]

    Kyle Akers and T. M. Murali. 2021. Gene regulatory network inference in single- cell biology.Current Opinion in Systems Biology26 (2021), 87–97. doi:10.1016/j. coisb.2021.04.007

    work page doi:10.1016/j 2021

  2. [2]

    Carmen Bravo González-Blas et al. 2023. SCENIC+: single-cell multiomic infer- ence of enhancers and gene regulatory networks.Nature Methods20, 9 (2023), 1355–1367. doi:10.1038/s41592-023-01938-4

    work page doi:10.1038/s41592-023-01938-4 2023

  3. [3]

    Zhi-Jie Cao and Ge Gao. 2022. Multi-omics single-cell data integration and regulatory inference with graph-linked embedding.Nature Biotechnology40, 10 (2022), 1458–1466. doi:10.1038/s41587-022-01284-4

    work page doi:10.1038/s41587-022-01284-4 2022

  4. [4]

    Shuonan Chen and Jessica C. Mar. 2018. Evaluating methods of inferring gene regulatory networks highlights their lack of performance for single cell gene 9 KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Tianyang et al. expression data.BMC Bioinformatics19, 1 (2018), 232. doi:10.1186/s12859-018- 2217-z

    work page doi:10.1186/s12859-018- 2018

  5. [5]

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. 2024. scGPT: toward building a foundation model for single- cell multi-omics using generative AI.Nature Methods21, 8 (2024), 1470–1480. doi:10.1038/s41592-024-02201-0

    work page doi:10.1038/s41592-024-02201-0 2024

  6. [6]

    Davidson

    Eric H. Davidson. 2006.The Regulatory Genome: Gene Regulatory Networks in Development and Evolution. Academic Press, Burlington, MA. https://www. sciencedirect.com/book/9780120885633/the-regulatory-genome

    work page arXiv 2006

  7. [7]

    Riet De Smet and Kathleen Marchal. 2010. Advantages and limitations of current network inference methods.Nature Reviews Microbiology8, 10 (2010), 717–729. doi:10.1038/nrmicro2419

    work page doi:10.1038/nrmicro2419 2010

  8. [8]

    Faith, Boris Hayete, Joshua T

    Jeremiah J. Faith, Boris Hayete, Joshua T. Thaden, Ilaria Mogno, Jeffrey Wierzbowski, Gilles Cottarel, Simon Kasif, James J. Collins, and Timothy S. Gard- ner. 2007. Large-Scale Mapping and Validation ofEscherichia coliTranscriptional Regulation from a Compendium of Expression Profiles.PLoS Biology5, 1 (2007), e8. doi:10.1371/journal.pbio.0050008

    work page doi:10.1371/journal.pbio.0050008 2007

  9. [9]

    Gray Camp, and Barbara Treutlein

    Jonas Simon Fleck, Sophie Martina Johanna Jansen, Damian Wollny, Fides Zenk, Makiko Seimiya, Akanksha Jain, Ryoko Okamoto, Malgorzata Santel, Zhisong He, J. Gray Camp, and Barbara Treutlein. 2023. Inferring and perturbing cell fate regulomes in human brain organoids.Nature621, 7978 (2023), 365–372. doi:10.1038/s41586-022-05279-8

    work page doi:10.1038/s41586-022-05279-8 2023

  10. [10]

    Holland, Mahmoud M

    Luz Garcia-Alonso, Christian H. Holland, Mahmoud M. Ibrahim, Denes Turei, and Julio Saez-Rodriguez. 2019. Benchmark and integration of resources for the estimation of human transcription factor activities.Genome Research29, 8 (2019), 1363–1375. doi:10.1101/gr.240663.118

    work page doi:10.1101/gr.240663.118 2019

  11. [11]

    Heonjong Han, Jae-Won Cho, Sangyoung Lee, Ayoung Yun, Hyojin Kim, Dasom Bae, Sunmo Yang, Chan Yeong Kim, Muyoung Lee, Eunbeen Lee, Sungho Lee, Byunghee Kang, Dabin Jeong, Yaeji Kim, Hyeon-Nae Jeon, Haein Jung, Sunhwee Nam, Michael Chung, Jong-Hoon Kim, and Insuk Lee. 2018. TRRUST v2: an expanded reference database of human and mouse transcriptional regula...

    work page doi:10.1093/nar/ 2018

  12. [12]

    Mauck III, Shiwei Zheng, Andrew Butler, Maddie J

    Yuhan Hao, Stephanie Hao, Erica Andersen-Nissen, William M. Mauck III, Shi- wei Zheng, Andrew Butler, Maddie J. Lee, Aleksander J. Wilk, Charlotte Darby, Michael Zager, Paul Hoffman, Marlon Stoeckius, Efthymia Papalexi, Eleni P. Mim- itou, Jay A. Jain, Avi Srivastava, et al. 2021. Integrated analysis of multimodal single-cell data.Cell184, 13 (2021), 3573...

    work page doi:10.1016/j.cell.2021.04.048 2021

  13. [13]

    Vân Anh Huynh-Thu, Alexandre Irrthum, Louis Wehenkel, and Pierre Geurts

  14. [14]

    doi:10.1371/journal.pone.0012776

    Inferring regulatory networks from expression data using tree-based methods.PLoS ONE5, 9 (2010), e12776. doi:10.1371/journal.pone.0012776

    work page doi:10.1371/journal.pone.0012776 2010

  15. [15]

    Guy Karlebach and Ron Shamir. 2008. Modelling and analysis of gene regulatory networks.Nature Reviews Molecular Cell Biology9, 10 (2008), 770–780. doi:10. 1038/nrm2503

    2008

  16. [16]

    Kharchenko, Lev Silberstein, and David T

    Peter V. Kharchenko, Lev Silberstein, and David T. Scadden. 2014. Bayesian approach to single-cell differential expression analysis.Nature Methods11, 7 (2014), 740–742. doi:10.1038/nmeth.2967

    work page doi:10.1038/nmeth.2967 2014

  17. [17]

    du Plessis, and Masashi Sugiyama

    Ryuichi Kiryo, Gang Niu, Marthinus C. du Plessis, and Masashi Sugiyama. 2017. Positive-unlabeled learning with non-negative risk estimator. InAdvances in Neural Information Processing Systems, Vol. 30. Curran Associates, Inc., Red Hook, NY, USA, 1675–1685

    2017

  18. [19]

    Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, Melissa Goldman, Itay Tirosh, Allison R

    Evan Z. Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, Melissa Goldman, Itay Tirosh, Allison R. Bialas, Nolan Kamitaki, Emily M. Marter- steck, John J. Trombetta, David A. Weitz, Joshua R. Sanes, Alex K. Shalek, Aviv Regev, and Steven A. McCarroll. 2015. Highly Parallel Genome-wide Expres- sion Profiling of Individual Cells Using Nan...

    work page doi:10.1016/j.cell.2015.05.002 2015

  19. [20]

    Margolin, Ilya Nemenman, Katia Basso, Chris Wiggins, Gustavo Stolovitzky, Riccardo Dalla Favera, and Andrea Califano

    Adam A. Margolin, Ilya Nemenman, Katia Basso, Chris Wiggins, Gustavo Stolovitzky, Riccardo Dalla Favera, and Andrea Califano. 2006. ARACNE: An Algorithm for the Reconstruction of Gene Regulatory Networks in a Mammalian Cellular Context.BMC Bioinformatics7, Suppl 1 (2006), S7. doi:10.1186/1471- 2105-7-S1-S7

    work page doi:10.1186/1471- 2006

  20. [21]

    Thomas Moerman, Sara Aibar, Carmen Bravo González-Blas, Jaak Simm, Yves Moreau, Jan Aerts, and Stein Aerts. 2019. GRNBoost2 and Arboreto: efficient and scalable inference of gene regulatory networks.Bioinformatics35, 12 (2019), 2159–2161. doi:10.1093/bioinformatics/bty916

    work page doi:10.1093/bioinformatics/bty916 2019

  21. [22]

    Ryan M Patrick, Rajeev Ranjan, Shantha R Sumanasinghe, Phillip M San Miguel, Kranthi Varala, and Ying Li. 2026. Every Cell Counts: Tomato Root Responses to Nitrogen at Single-Cell Resolution.bioRxiv(2026). arXiv:https://www.biorxiv.org/content/early/2026/02/07/2026.02.06.704465.full.pdf doi:10.64898/2026.02.06.704465

    work page doi:10.64898/2026.02.06.704465 2026

  22. [23]

    Griffiths, Caroline Guibentif, Thomas W

    Blanca Pijuan-Sala, Jonathan A. Griffiths, Caroline Guibentif, Thomas W. Hiscock, Wardha Jawaid, Fernando J. Calero-Nieto, Carla Mulas, Ximena Ibarra-Soria, Richard C. V. Tyser, Dominic L. L. Ho, Wolf Reik, Shankar Srinivas, Benjamin D. Simons, Jennifer Nichols, John C. Marioni, and Berthold Gottgens. 2019. A single- cell molecular map of mouse gastrulati...

    work page doi:10.1038/s41586-019-0933-9 2019

  23. [24]

    Pliner, Jonathan S

    Samantha A. Pliner, Jonathan S. Packer, José L. McFaline-Figueroa, Darren A. Cusanovich, Riza M. Daza, Ilya Aigha, Charles P. Fulco, Jason H. Bielas, Sebastian Preissl, Grace X. Y. Zheng, and Cole Trapnell. 2018. Cicero predicts cis-regulatory DNA interactions from single-cell chromatin accessibility data.Molecular Cell 71, 5 (2018), 858–871.e8. doi:10.10...

    work page doi:10.1016/j.molcel.2018.06.044 2018

  24. [25]

    Jalihal, Jeffrey N

    Aditya Pratapa, Amogh P. Jalihal, Jeffrey N. Law, Aditya Bharadwaj, and T. M. Murali. 2020. Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data.Nature Methods17, 2 (2020), 147–154. doi:10. 1038/s41592-019-0690-6

    2020

  25. [26]

    Schep, Botao Wu, Jason D

    Anthony N. Schep, Botao Wu, Jason D. Buenrostro, and William J. Greenleaf. 2017. chromVAR: inferring transcription-factor-associated accessibility from single-cell epigenomic data.Nature Methods14, 10 (2017), 975–978. doi:10.1038/nmeth.4401

    work page doi:10.1038/nmeth.4401 2017

  26. [27]

    Stefan Schoenfelder and Peter Fraser. 2019. Long-range enhancer–promoter contacts in gene expression control.Nature Reviews Genetics20, 8 (2019), 437–

    2019

  27. [28]

    doi:10.1038/s41576-019-0128-0

    work page doi:10.1038/s41576-019-0128-0

  28. [29]

    Marco Stock, Corinna Losert, Matteo Zambon, Niclas Popp, Gabriele Lubatti, Eva Hörmanseder, Matthias Heinig, and Antonio Scialdone. 2025. Leveraging prior knowledge to infer gene regulatory networks from single-cell RNA-sequencing data.Molecular Systems Biology21, 3 (2025), 214–230. doi:10.1038/s44320-025- 00088-3

    work page doi:10.1038/s44320-025- 2025

  29. [30]

    Mauck, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija

    Tim Stuart, Andrew Butler, Paul Hoffman, Christoph Hafemeister, Efthymia Papalexi, William M. Mauck III, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija. 2019. Comprehensive integration of single-cell data.Cell177, 7 (2019), 1888–1902.e21. doi:10.1016/j.cell.2019.05.031

    work page doi:10.1016/j.cell.2019.05.031 2019

  30. [31]

    Theodoris, Ling Xiao, Anant Chopra, Mark D

    Christina V. Theodoris, Ling Xiao, Anant Chopra, Mark D. Chaffin, Zeina R. Al Sayed, Matthew C. Hill, Helene Mantineo, Elizabeth M. Brydon, Zexian Zeng, X. Shirley Liu, and Patrick T. Ellinor. 2023. Transfer learning enables predictions in network biology.Nature618, 7965 (2023), 616–624. doi:10.1038/s41586-023- 06139-9

    work page doi:10.1038/s41586-023- 2023

  31. [32]

    Lennon, Kenneth J

    Cole Trapnell, Davide Cacchiarelli, Justin Grimsby, Pallavi Pokharel, Shuqiang Li, Michael Morse, Niall J. Lennon, Kenneth J. Livak, Tarjei S. Mikkelsen, and John L. Rinn. 2014. The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells.Nature Biotechnology32 (2014), 381–386. doi:10.1038/nbt.2859

    work page doi:10.1038/nbt.2859 2014

  32. [33]

    Grace X. Y. Zheng, Julie M. Terry, Phillip Belgrader, Paul Ryvkin, Zachary W. Bent, Robert Wilson, Sean B. Ziraldo, Tobias D. Wheeler, Geoff P. McDermott, Junjie Zhu, Michael T. Gregory, Joe Shuga, Lydia Montesclaros, James G. Under- wood, Donald A. Masquelier, Sara Y. Nishimura, Michael Schnall-Levin, Peter W. Wyatt, Colin M. Hindson, Rajiv Bharadwaj, Al...

    work page doi:10.1038/ncomms14049 2017