Pith. sign in

REVIEW 3 major objections 8 minor 55 references

This paper claims that GRN inference is better framed as an inductive, ranking-based graph completion task, and that a joint discrete diffusion over gene states and edges achieves new state-of-the-art results on the proposed BEELINE-KGC ben

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 06:10 UTC pith:RWU67EOQ

load-bearing objection The BEELINE-KGC evaluation protocol is a real contribution; the co-evolutionary diffusion at the core of CoDiffGRN is not well-defined as written. the 3 major comments →

arxiv 2607.13120 v1 pith:RWU67EOQ submitted 2026-07-14 cs.LG cs.AI

CoDiffGRN: Rethinking Gene Regulatory Network Inference via the BEELINE-KGC Benchmark and Co-evolutionary Discrete Diffusion

Jiaze Song , Runhao Zhao , Minghao Xu , Bin Cui , Wentao Zhang This is my paper
classification cs.LG cs.AI
keywords gene regulatory network inferenceknowledge graph completioninductive gene-holdout splitdiscrete diffusionco-evolutionary diffusiontop-K rankingsingle-cell RNA-seqBEELINE-KGC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard benchmarks for gene regulatory network (GRN) inference judge models on all gene pairs at once using global metrics, but biologists actually want a shortlist of high-confidence regulators for novel, previously unseen genes. The paper tries to fix this at both ends: it defines a new evaluation protocol, BEELINE-KGC, that holds out transcription factors (TFs) during training and scores only top-ranked predictions with knowledge-graph-completion metrics, and it introduces CoDiffGRN, a diffusion model that co-evolves discretized gene expression states and regulatory edges so that edge transitions depend on the activity states of their endpoint genes. The authors claim that CoDiffGRN ranks first on every cell type and metric in both the TFs+500 and TFs+1000 settings of the new benchmark, with average relative gains of 24.5% and gains up to 184.8% over the strongest baselines on the Specific network. If this holds, it suggests that inductive, ranking-centric evaluation together with genuinely conditional generative modeling can produce substantially more reliable predictions of novel regulatory interactions than existing GRN inference methods.

Core claim

The paper's central claim is that prior GRN inference fails in the setting that matters for real discovery — predicting targets of transcription factors that were never seen in training — because benchmarks hide this failure behind transductive splits and global AUROC/AUPRC metrics. To expose it, the authors recast GRN inference as a knowledge-graph completion problem with an inductive gene-holdout split, scoring Hits@K and MRR on held-out TFs. To solve it, they propose CoDiffGRN: a joint node-edge discrete diffusion process in which discretized gene expression states and edge existence co-evolve, with a transition matrix whose cross-terms couple edge changes to the states of endpoint genes,

What carries the argument

The load-bearing mechanism is the joint discrete diffusion over a combined gene-edge state space. A single transition matrix QG is built from four blocks: gene-state transitions (QD), edge transitions (QE), and cross-transitions between gene states and edge types (QDE, QED), so that an edge's noise/denoise behavior is conditioned on the discretized activity of its two endpoint genes. This is paired with cell-cluster discretization, which turns continuous expression profiles into k-bit activation patterns (one-hot over 2^k states) to give unseen genes a stable distributional anchor, and with TF-ALL Subgraph Sampling (TASS), which samples TF-enriched subgraphs for scalable training and average

Load-bearing premise

The load-bearing premise is that the inductive gene-holdout split—where entire TFs and their edges are withheld during training—faithfully simulates the conditions of real regulatory discovery, and that top-K ranking accuracy under that split will translate into successful wet-lab validation.

What would settle it

Run a prospective validation: take CoDiffGRN's top-10 target predictions for held-out TFs in a new cell type, test them with ChIP-seq or perturbation assays, and measure the confirmation rate; if it is no higher than the confirmation rate of the strongest baseline (or of random TF–target candidates), the claim that CoDiffGRN improves novel regulatory discovery is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On the BEELINE-KGC protocol, CoDiffGRN's Hits@10 and Hits@50 are consistently above every baseline for all seven cell types in both TFs+500 and TFs+1000 settings, so a biologist using its top-K list could expect more true targets per validation batch.
  • Because the benchmark scores held-out TFs, models that win on it are demonstrably better at generalizing to unseen regulators, not just at reconstructing known edges.
  • The ablation results imply that both the gene-edge co-evolution and the task-aware subgraph sampling contribute, meaning future GRN methods should condition edge transitions on gene state rather than applying uniform noise.
  • The paper's ablations show that both components matter; without joint modeling or with uniform random subgraph sampling, performance degrades, indicating the gains are not from the backbone alone.

Where Pith is reading between the lines

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

  • Even at the new state of the art, absolute Hits@10 values stay low (0.016–0.334), so the practical ceiling on novel-TF discovery remains modest; a biologist would still need many candidate validations per hit, and the relative 184.8% gain is over a very small absolute base.
  • The inductive gene-holdout split assumes held-out TFs behave like truly novel TFs; if unknown regulators interact through different mechanisms than known ones, the benchmark could overstate or understate real-world transfer. A direct test would be to compare model rankings against a prospective ChIP-seq or CRISPR screen on a cell type not in the training set.
  • The co-evolutionary discrete diffusion idea is transferable to other graph-completion problems where node states and edge existence are mutually dependent, such as protein-protein interaction networks or drug-target graphs, though the paper does not test this.
  • The TASS subgraph sampling could be made adaptive or theory-driven; the paper sets m to a rounded theoretical bound, but the effect of m and subgraph size k on ranking stability is not swept, so a sensitivity analysis would clarify how much of the gain is from sampling versus modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. The paper argues that standard GRN inference evaluation is misaligned with practical biological discovery: benchmarks use transductive splits and global metrics (AUROC/AUPRC), whereas biologists need inductive generalization to unseen TFs and top-K ranking quality for experimental validation. To address this, the authors introduce BEELINE-KGC, a benchmark with an inductive gene-holdout split and knowledge-graph-completion metrics (Hits@K, MRR). They also propose CoDiffGRN, a discrete diffusion model that jointly models discretized gene expression states and regulatory edges via a 'co-evolutionary' transition matrix, together with a TF-ALL Subgraph Sampling (TASS) strategy for scalable training. Experiments on seven cell types report consistent improvements over existing GNN-, diffusion-, and feature-based methods, and ablations support the contributions of the joint diffusion and TASS.

Significance. If the technical claims hold, the BEELINE-KGC benchmark is a timely and practically relevant contribution because it reframes GRN inference evaluation around the inductive, ranking-centric setting that experimental biologists face. CoDiffGRN's reported consistent gains over diverse baselines are encouraging and suggest that joint node-edge discrete diffusion and subgraph augmentation may help in this setting. However, the mathematical formulation of the joint diffusion is not valid as written (Eq. 7), and key experimental details are underspecified, so the state-of-the-art claim cannot currently be taken at face value. The paper also reports only means without variance, making the strength of the empirical conclusions uncertain. The work would benefit from a corrected formalism, a reproducibility package, and a more careful discussion of absolute performance levels.

major comments (3)
  1. [Section 4.2, Eq. (7)] The joint transition matrix in Eq. (7) is not a transition matrix on the state space defined in Section 4.2. The node token G[i] concatenates a one-hot node state in R^{F_D} with N one-hot edge states in R^2; the joint configuration space has F_D * 2^N states, requiring a transition matrix of size (F_D * 2^N) x (F_D * 2^N). Eq. (7) instead has dimension (F_D + 2N) x (F_D + 2N), which is a transition matrix for a single categorical variable with F_D+2N states, not for the joint node-edge state. The off-diagonal blocks Q_DE and Q_ED also allow probability mass to flow directly between gene-state categories and edge-presence categories, which is not a valid transition in the product space. As written, the forward process q(G_t | G_{t-1}) is not well-defined. If the actual implementation uses separate transition matrices for node and edge types (as in DiGress, Ref. [38]), the authors must re
  2. [Section 4.4, Algorithm 2, Section 5.1, Table 2] The reported SOTA numbers are not reproducible from the manuscript. (i) In Section 4.4, the number of subgraphs is stated as m = ceil(p^{-2} log n log(1/delta)), but the confidence parameter delta is never specified; Algorithm 2 uses the same undefined quantity. (ii) The inductive gene-holdout split (Section 3.2) does not describe how V_hold is chosen (random seed, degree-based, etc.), yet the results in Table 2 depend on this choice. (iii) Section 5.1 says 'with both mean and standard deviation reported,' but Table 2 and Figure 3 report only means with no variance or error bars. (iv) Code is not provided. These gaps make it impossible to verify the central claim that CoDiffGRN 'ranks first across all cell types and metrics in both TFs+500 and TFs+1000 settings.' At minimum, the authors should specify all hyperparameters (including delta), the split-selection protocol, and release code a
  3. [Table 2, Figure 3] The practical motivation in Section 3.2 hinges on the usefulness of top-K predictions for experimental validation. However, absolute Hits@10 values in Table 2 are low: for CoDiffGRN they range from .016 to .334 across datasets, with most values below .1. This means that for a large fraction of holdout TF queries, no true target appears in the top 10. While relative improvements over baselines are consistent, the paper should explicitly discuss whether such absolute performance is sufficient for the claimed 'reliable biological discovery' use case. Additionally, Figure 3 presents ablations without error bars or significance tests; the reported differences among Edge Only, w/o TASS, and RandSS may not be statistically reliable. Adding variance estimates and, where feasible, statistical tests would strengthen the conclusions.
minor comments (8)
  1. [Section 4.1, Eq. (6)] Eq. (6) appears to mislabel the posterior distribution as the joint. The left-hand side should be q(x_{t-1} | x_t, x_0), not q(x_{t-1}, x_t | x_0). The displayed expression is the standard D3PM posterior, and the current notation is confusing.
  2. [Section 4.2] The definition of the global graph token G[i] is unclear: it is defined per node, but later the text refers to a 'global graph token' as if it were a single object. Please clarify whether the forward process is applied independently to each node's token or to the entire graph as a product space.
  3. [Section 5.1] The paper states that both mean and standard deviation are reported, but Table 2 and Figure 3 show only means. Please either include standard deviations or revise the text.
  4. [Algorithm 1] The term X_vis is not defined. Presumably this is the expression matrix restricted to the visible (training) genes, but it should be stated explicitly.
  5. [Table 1] Table 1's formatting is dense and difficult to read, with values such as '34 (34) 815 (1260) 3535 (5995)' running together. Please reformat for clarity.
  6. [Reference [14]] The author list in reference [14] contains a typo: 'andT M Murali' should be 'and T. M. Murali'.
  7. [Introduction / Related Work] The paper claims 'the first co-evolutionary discrete diffusion framework' for GRNs, but Ref. [37] already uses a joint node-edge discrete diffusion. The authors should more carefully distinguish their contribution from that work.
  8. [Section 4.4] The rounding 'm = 100 * ceil(m_base / 100)' is arbitrary; please justify or at least explain the choice of the factor 100.

Circularity Check

0 steps flagged

No significant circularity: CoDiffGRN's reported gains are not forced by its own construction; training and evaluation are separated by the inductive holdout, key building blocks are cited externally, and transductive results on the standard BEELINE setting provide an independent check.

full rationale

I walked the derivation chain from the benchmark formalization to the model's reported results. The inductive gene-holdout split withholds all edges involving holdout TFs from training; the model is trained only on visible subgraphs, and the held-out labels enter only in the final Hits@K/MRR computation. The discretization (Algorithm 1) explicitly constructs clusters from `X_vis` (visible data), and the D3PM-style transition matrices `Q_D`/`Q_E` are derived from training-data marginals following external references (D3PM [18], DiGress [38]), not from the test labels. The cross-dependency matrices `Q_ED`/`Q_DE` are a modeling choice stated in the paper, not a fitted parameter renamed as a prediction. There are no load-bearing self-citations: all key methodological citations (D3PM, DiGress, GATv2, SAGESS) are external, and no author-owned uniqueness theorem is invoked to force the model choice. The appendix reports results on the standard transductive BEELINE setting, providing an external-benchmark check that the SOTA claim is not solely an artifact of the newly introduced BEELINE-KGC protocol. The potential dimensional inconsistency of Eq. 7 is a formal correctness concern about whether the joint transition matrix is well-defined, but it is not a case where a prediction is equivalent to its input by construction, so it does not constitute circularity under the rules of this pass. The limitations section mentions restricted biological contexts and computational overhead, but does not admit or rely on a circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper does not introduce new physical or biological entities (no new genes, forces, or particles). Its 'inventions' are conceptual/algorithmic: the BEELINE-KGC protocol and the CoDiffGRN architecture. The most heavily weighted assumptions are the validity of the BEELINE ground truths and the faithfulness of the inductive split. Free parameters are mostly standard hyperparameters, but note that the discretization threshold and TASS sampling fraction are data-derived and enter the central result.

free parameters (5)
  • Number of cell clusters k = 4
    Chosen to yield 16 discrete states; no sensitivity analysis is reported. The discretization granularity is a modeling choice that controls the resulting state space (Sec. 4.2, Sec. 5.1).
  • Cell-cluster activation threshold tau_j = computed from visible data (Algorithm 1)
    The binarization threshold is derived from the mean expression of visible data within each cluster. This is a data-derived threshold, equivalent to a fit to the training distribution.
  • TASS subgraph count m (and sampling ratio t) = m = 100 * ceil(m_base/100), t = |V_TF|/|V_G|
    The sampling ratio and rounding to nearest multiple of 100 are implementation choices that affect the training distribution and inference-time aggregation (Algorithm 2).
  • Diffusion steps = 500
    Chosen hyperparameter; no ablation on the number of steps (Sec. 5.1).
  • lambda auxiliary loss weight = lambda << 1
    The strength of the node-level MSE auxiliary objective is stated only as lambda << 1; exact value not specified (Eq. 11).
axioms (5)
  • domain assumption BEELINE ground-truth networks (Specific, Non-Specific, STRING, LOF/GOF) are a valid proxy for true regulatory interactions.
    The benchmark's evaluation is only as meaningful as the ground-truth networks it uses. The paper cites BEELINE [14] and the underlying databases (ENCODE, ChIP-Atlas, DoRothEA, STRING, etc.) as truth, but these are known to be incomplete and noisy.
  • domain assumption Cell-cluster-based binarization into k activation bits preserves enough biological signal for diffusion to be meaningful.
    Introduced in Sec. 4.2 and motivated in Appendix B; the appendix supports it with UMAP and activation consistency figures, not with a quantitative downstream task.
  • standard math D3PM/categorical diffusion framework is applicable to GRN topology and gene states.
    The paper builds on D3PM [18] and DiGress [38]; this is a prior-art framework assumption.
  • domain assumption The inductive gene-holdout split, generated from the BEELINE graphs by the authors, adequately approximates real-world novel-TF discovery.
    This is the central premise of the benchmark. Real novel-TF discovery may involve genes with very different expression distributions, whereas this split is a random holdout of TFs from the same graph/dataset.
  • domain assumption The filtered KGC ranking (excluding other known valid targets) is a faithful model of the experimental-validation budget.
    Filtered ranking is standard in KGC, but in GRN discovery, a biologist would not typically know all other valid targets and the 'filtered' setting may inflate the perceived quality of the top-K list.

pith-pipeline@v1.3.0-alltime-deepseek · 17441 in / 8393 out tokens · 154788 ms · 2026-08-02T06:10:51.295722+00:00 · methodology

0 comments
read the original abstract

Inferring gene regulatory networks (GRNs) from single-cell transcriptomic data is crucial for biological discovery, yet existing approaches suffer from a fundamental misalignment with real-world needs. Researchers typically seek a small set of high-confidence regulatory interactions for experimental validation, often involving previously unseen genes. However, current benchmarks rely on transductive splits with global classification metrics, while prevailing models struggle to generalize under inductive settings. To bridge this gap, we reformulate GRN inference as an inductive, ranking-centric graph completion problem and introduce \textbf{\benchmark}, a new benchmark that incorporates an inductive gene-holdout split together with knowledge graph completion metrics to better evaluate top-ranked predictions. Building on this, we propose \textbf{\method}, the first co-evolutionary discrete diffusion framework that jointly models biologically coherent discretized gene expression states and regulatory interactions for robust inductive generalization and improved top-ranked regulatory discovery. We further introduce TF-ALL Subgraph Sampling (TASS) for scalable training. Extensive experiments on {\benchmark} show that {\method} establishes new state-of-the-art performance, significantly outperforming existing methods in novel regulatory discovery, and ablation studies further verify the effectiveness of our design.

Figures

Figures reproduced from arXiv: 2607.13120 by Bin Cui, Jiaze Song, Minghao Xu, Runhao Zhao, Wentao Zhang.

Figure 1
Figure 1. Figure 1: Overview of current limitations and our proposed solutions. (a) Current benchmarks mismatch with realistic demand of inductive generalization and top-K ranking quality, while existing methods degrade with unseen genes and fixed edge transitions. (b) Our BEELINE-KGC introduces an inductive, ranking-based evaluation protocol, and CoDiffGRN leverages cell-cluster discretization, co-evolutionary discrete diffu… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the CoDiffGRN framework. The framework combines TF-ALL Subgraph Sampling (TASS) with a co-evolutionary discrete diffusion model. TASS enables efficient training of a multi-conditioned TF-Aware Graph Denoising Module. At inference, predictions from multiple subgraphs are aggregated via the reverse diffusion process. Abbr., Aggr.: Aggregation. Cell-cluster discretization. To mitigate feature coll… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study results on BEELINE-KGC (Specific network). We report mean performance of Hits@10 and Hits@50 over runs, under the TFs+500 (top) and TFs+1000 (bottom) settings. Abbr., w/o: without. • GNN-based models show limited inductive generalization. While GENELink+ remains com￾petitive, most GNN methods suffer from low performance. Their heavy reliance on predefined prior graphs leads to disconnected c… view at source ↗
Figure 4
Figure 4. Figure 4: Cell-state structure induced by clustering. Visualization of cell clusters obtained from scRNA-seq profiles using UMAP [41]. Each cluster corresponds to a distinct cellular state, supporting the use of cell-level structure as the basis for discretization. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Gene activation consistency across cell clusters. Representative genes show stable and coherent activation patterns within each cell cluster, indicating that discretization preserves biologically meaningful gene activity structure. Design Principle. Our method explicitly leverages this structure by first clustering cells into k biologically meaningful groups, and then constructing gene-level discrete repre… view at source ↗
Figure 6
Figure 6. Figure 6: Discrete gene manifold structure. The induced gene representation space exhibits smooth and coherent structure, indicating that the discretization preserves biologically meaningful organization while enabling stable diffusion over discrete states. Together, these results support the role of Cell-cluster-based Discretization as a biologically grounded representation scheme that enables stable inductive gene… view at source ↗

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

55 extracted references · 3 linked inside Pith

  1. [1]

    Protein post-translational modi- fications and regulation of pluripotency in human stem cells.Cell research, 24(2):143–160, 2014

    Yu-Chieh Wang, Suzanne E Peterson, and Jeanne F Loring. Protein post-translational modi- fications and regulation of pluripotency in human stem cells.Cell research, 24(2):143–160, 2014

    2014

  2. [2]

    Cooperative binding of transcription factors orchestrates reprogramming.Cell, 168(3):442–459, 2017

    Constantinos Chronis, Petko Fiziev, Bernadett Papp, Stefan Butz, Giancarlo Bonora, Shan Sabri, Jason Ernst, and Kathrin Plath. Cooperative binding of transcription factors orchestrates reprogramming.Cell, 168(3):442–459, 2017

    2017

  3. [3]

    Gene regulatory network inference: data integration in dynamic models—a review

    Michael Hecker, Sandro Lambeck, Susanne Toepfer, Eugene Van Someren, and Reinhard Guthke. Gene regulatory network inference: data integration in dynamic models—a review. Biosystems, 96(1):86–103, 2009

    2009

  4. [4]

    Cis-regulatory elements: molecular mechanisms and evolutionary processes underlying divergence.Nature Reviews Genetics, 13(1):59–69, 2012

    Patricia J Wittkopp and Gizem Kalay. Cis-regulatory elements: molecular mechanisms and evolutionary processes underlying divergence.Nature Reviews Genetics, 13(1):59–69, 2012

    2012

  5. [5]

    The technology and biology of single-cell rna sequencing.Molecular cell, 58(4):610–620, 2015

    Aleksandra A Kolodziejczyk, Jong Kyoung Kim, Valentine Svensson, John C Marioni, and Sarah A Teichmann. The technology and biology of single-cell rna sequencing.Molecular cell, 58(4):610–620, 2015

    2015

  6. [6]

    Revealing the vectors of cellular identity with single-cell genomics.Nature biotechnology, 34(11):1145–1160, 2016

    Allon Wagner, Aviv Regev, and Nir Yosef. Revealing the vectors of cellular identity with single-cell genomics.Nature biotechnology, 34(11):1145–1160, 2016

    2016

  7. [7]

    Bayesian approach to single-cell differential expression analysis.Nature methods, 11(7):740–742, 2014

    Peter V Kharchenko, Lev Silberstein, and David T Scadden. Bayesian approach to single-cell differential expression analysis.Nature methods, 11(7):740–742, 2014

    2014

  8. [8]

    Graph attention network for link prediction of gene regulations from single-cell rna-sequencing data.Bioinformatics, 38(19):4522–4529, 2022

    Guangyi Chen and Zhi-Ping Liu. Graph attention network for link prediction of gene regulations from single-cell rna-sequencing data.Bioinformatics, 38(19):4522–4529, 2022

    2022

  9. [9]

    Predicting gene regulatory links from single-cell rna-seq data using graph neural networks

    Guo Mao, Zhengbin Pang, Ke Zuo, Qinglin Wang, Xiangdong Pei, Xinhai Chen, and Jie Liu. Predicting gene regulatory links from single-cell rna-seq data using graph neural networks. Briefings in Bioinformatics, 24(6):bbad414, 2023

    2023

  10. [10]

    Inferring cell-type-specific gene regulatory network from cellular transcriptomics data with genelink+.Briefings in Bioinformatics, 26(4):bbaf359, 2025

    Wei Zhang, Bowen Shao, Wenrui Li, Wenbo Guo, Jiaxin Lyu, Guangyi Chen, Chuanyuan Wang, and Zhi-Ping Liu. Inferring cell-type-specific gene regulatory network from cellular transcriptomics data with genelink+.Briefings in Bioinformatics, 26(4):bbaf359, 2025

    2025

  11. [11]

    Grnformer: accurate gene regulatory network inference using graph transformer.Bioinformatics, 42(4):btag144, 2026

    Akshata Hegde and Jianlin Cheng. Grnformer: accurate gene regulatory network inference using graph transformer.Bioinformatics, 42(4):btag144, 2026

    2026

  12. [12]

    Gclink: a graph contrastive link prediction framework for gene regulatory network inference.Bioinformatics, 41(3):btaf074, 2025

    Weiming Yu, Zerun Lin, Miaofang Lan, and Le Ou-Yang. Gclink: a graph contrastive link prediction framework for gene regulatory network inference.Bioinformatics, 41(3):btaf074, 2025

    2025

  13. [13]

    Inferring gene regulatory network from single-cell transcriptomes with graph autoencoder model.PLoS Genetics, 19(9):e1010942, 2023

    Jiacheng Wang, Yaojia Chen, and Quan Zou. Inferring gene regulatory network from single-cell transcriptomes with graph autoencoder model.PLoS Genetics, 19(9):e1010942, 2023

    2023

  14. [14]

    Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data.Nature methods, 17(2):147–154, 2020

    Aditya Pratapa, Amogh P Jalihal, Jeffrey N Law, Aditya Bharadwaj, and andT M Murali. Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data.Nature methods, 17(2):147–154, 2020

    2020

  15. [15]

    Dissecting gene regulatory networks governing human cortical cell fate.Nature, pages 1–11, 2026

    Jingwen W Ding, Chang N Kim, Megan S Ostrowski, Yashodara Abeykoon, Bryan J Pavlovic, Jenelle L Wallace, Nathan K Schaefer, Tomasz J Nowakowski, and Alex A Pollen. Dissecting gene regulatory networks governing human cortical cell fate.Nature, pages 1–11, 2026

    2026

  16. [16]

    Identifying key underlying regulatory networks and predicting targets of orphan c/d box snord116 snornas in prader–willi syndrome.Nucleic Acids Research, 52(22): 13757–13774, 2024

    Rachel B Gilmore, Yaling Liu, Christopher E Stoddard, Michael S Chung, Gordon G Carmichael, and Justin Cotney. Identifying key underlying regulatory networks and predicting targets of orphan c/d box snord116 snornas in prader–willi syndrome.Nucleic Acids Research, 52(22): 13757–13774, 2024

    2024

  17. [17]

    Translating embeddings for modeling multi-relational data.Advances in neural information processing systems, 26, 2013

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data.Advances in neural information processing systems, 26, 2013. 11

    2013

  18. [18]

    Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    2021

  19. [19]

    Benchmarking frameworks for gene regulatory network inference methods: A review.NIPES JSTR SPECIAL ISSUE, 7(1):2224–2228, 2025

    Itunuoluwa Isewon, Mariam Kanonte, Emmanuel Alagbe, and Jelili Oyelade. Benchmarking frameworks for gene regulatory network inference methods: A review.NIPES JSTR SPECIAL ISSUE, 7(1):2224–2228, 2025

    2025

  20. [20]

    Grnboost2 and arboreto: efficient and scalable inference of gene regulatory networks.Bioinformatics, 35(12):2159–2161, 2019

    Thomas Moerman, Sara Aibar Santos, Carmen Bravo González-Blas, Jaak Simm, Yves Moreau, Jan Aerts, and Stein Aerts. Grnboost2 and arboreto: efficient and scalable inference of gene regulatory networks.Bioinformatics, 35(12):2159–2161, 2019

    2019

  21. [21]

    Deep learning for inferring gene relationships from single-cell expression data.Proceedings of the National Academy of Sciences, 116(52):27151–27158, 2019

    Ye Yuan and Ziv Bar-Joseph. Deep learning for inferring gene relationships from single-cell expression data.Proceedings of the National Academy of Sciences, 116(52):27151–27158, 2019

    2019

  22. [22]

    Inferring regulatory networks from expression data using tree-based methods.PloS one, 5(9):e12776, 2010

    Vân Anh Huynh-Thu, Alexandre Irrthum, Louis Wehenkel, and Pierre Geurts. Inferring regulatory networks from expression data using tree-based methods.PloS one, 5(9):e12776, 2010

    2010

  23. [23]

    Modeling gene regulatory networks using neural network architectures.Nature Computational Science, 1(7):491–501, 2021

    Hantao Shu, Jingtian Zhou, Qiuyu Lian, Han Li, Dan Zhao, Jianyang Zeng, and Jianzhu Ma. Modeling gene regulatory networks using neural network architectures.Nature Computational Science, 1(7):491–501, 2021

    2021

  24. [24]

    From noise to knowledge: diffusion probabilistic model-based neural inference of gene regulatory networks.Journal of Computational Biology, 31(11): 1087–1103, 2024

    Hao Zhu and Donna Slonim. From noise to knowledge: diffusion probabilistic model-based neural inference of gene regulatory networks.Journal of Computational Biology, 31(11): 1087–1103, 2024

    2024

  25. [25]

    Diffusion-based generation of gene regulatory networks from scrna-seq data with dignet.Genome Research, 35(2):340–354, 2025

    Chuanyuan Wang and Zhi-Ping Liu. Diffusion-based generation of gene regulatory networks from scrna-seq data with dignet.Genome Research, 35(2):340–354, 2025

    2025

  26. [26]

    Attention-guided probabilistic diffusion model for generating cell-type-specific gene regulatory networks from gene expression profiles.Genes, 16(11):1255, 2025

    Shiyu Xu, Na Yu, Daoliang Zhang, and Chuanyuan Wang. Attention-guided probabilistic diffusion model for generating cell-type-specific gene regulatory networks from gene expression profiles.Genes, 16(11):1255, 2025

    2025

  27. [27]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  28. [28]

    Genome- wide transcription factor–binding maps reveal cell-specific changes in the regulatory architecture of human hspcs.Blood, 142(17):1448–1462, 2023

    Shruthi Subramanian, Julie AI Thoms, Yizhou Huang, Paola Cornejo-Páramo, Forrest C Koch, Sebastien Jacquelin, Sylvie Shen, Emma Song, Swapna Joshi, Chris Brownlee, et al. Genome- wide transcription factor–binding maps reveal cell-specific changes in the regulatory architecture of human hspcs.Blood, 142(17):1448–1462, 2023

    2023

  29. [29]

    Genome-wide crispr screen identifies irf1 and tfap4 as transcriptional regulators of galectin-9 in t cell acute lymphoblastic leukemia.Science advances, 11(12):eads8351, 2025

    Caroline RM Wiggers, Burak Yüzügüldü, Nathanial G Tadros, Tayla B Heavican-Foral, Eu- gene Y Cho, Zachary C Eisenbies, Merve Ozdemir, Steffen B Kulp, Yun-Cheol Chae, Alejandro Gutierrez, et al. Genome-wide crispr screen identifies irf1 and tfap4 as transcriptional regulators of galectin-9 in t cell acute lymphoblastic leukemia.Science advances, 11(12):ead...

    2025

  30. [30]

    htftarget: a comprehensive database for regulations of human transcription factors and their targets.Genomics, proteomics & bioinformatics, 18(2):120–128, 2020

    Qiong Zhang, Wei Liu, Hong-Mei Zhang, Gui-Yan Xie, Ya-Ru Miao, Mengxuan Xia, and An-Yuan Guo. htftarget: a comprehensive database for regulations of human transcription factors and their targets.Genomics, proteomics & bioinformatics, 18(2):120–128, 2020

    2020

  31. [31]

    Genome-wide (chip-seq) identification of target genes regulated by wrky33 during submergence stress in arabidopsis.BMC genomic data, 22(1):16, 2021

    Junlin Zhang, Bao Liu, Yan Song, Yang Chen, Jiao Fu, Jianquan Liu, Tao Ma, Zhenxiang Xi, and Huanhuan Liu. Genome-wide (chip-seq) identification of target genes regulated by wrky33 during submergence stress in arabidopsis.BMC genomic data, 22(1):16, 2021

    2021

  32. [32]

    Diverse transcription factor binding features revealed by genome-wide chip-seq in c

    Wei Niu, Zhi John Lu, Mei Zhong, Mihail Sarov, John I Murray, Cathleen M Brdlik, Judith Janette, Chao Chen, Pedro Alves, Elicia Preston, et al. Diverse transcription factor binding features revealed by genome-wide chip-seq in c. elegans.Genome research, 21(2):245–254, 2011

    2011

  33. [33]

    Molecular circuit between aspergillus nidulans transcription factors msna and velb to coordinate fungal stress and developmental responses.PLoS Genetics, 21(7):e1011578, 2025

    Emmanouil Bastakis, Jennifer Gerke, Seyma Özkan, Rebekka Harting, Tanja Lienard, Christoph Sasse, Emmanouil S Xylakis, Merle Aden, Anja Strohdiek, Gabriele Heinrich, et al. Molecular circuit between aspergillus nidulans transcription factors msna and velb to coordinate fungal stress and developmental responses.PLoS Genetics, 21(7):e1011578, 2025. 12

    2025

  34. [34]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  35. [35]

    Single-cell messenger rna sequencing reveals rare intestinal cell types.Nature, 525(7568):251–255, 2015

    Dominic Grün, Anna Lyubimova, Lennart Kester, Kay Wiebrands, Onur Basak, Nobuo Sasaki, Hans Clevers, and Alexander Van Oudenaarden. Single-cell messenger rna sequencing reveals rare intestinal cell types.Nature, 525(7568):251–255, 2015

    2015

  36. [36]

    Differential analysis of binarized single-cell rna sequencing data captures biological variation.NAR genomics and bioinformatics, 3(4):lqab118, 2021

    Gerard A Bouland, Ahmed Mahfouz, and Marcel JT Reinders. Differential analysis of binarized single-cell rna sequencing data captures biological variation.NAR genomics and bioinformatics, 3(4):lqab118, 2021

    2021

  37. [37]

    Graph diffusion transformers for multi- conditional molecular generation.Advances in Neural Information Processing Systems, 37: 8065–8092, 2024

    Gang Liu, Jiaxin Xu, Tengfei Luo, and Meng Jiang. Graph diffusion transformers for multi- conditional molecular generation.Advances in Neural Information Processing Systems, 37: 8065–8092, 2024

    2024

  38. [38]

    Digress: Discrete denoising diffusion for graph generation.arXiv preprint arXiv:2209.14734, 2022

    Clement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, V olkan Cevher, and Pas- cal Frossard. Digress: Discrete denoising diffusion for graph generation.arXiv preprint arXiv:2209.14734, 2022

    Pith/arXiv arXiv 2022

  39. [39]

    How attentive are graph attention networks?arXiv preprint arXiv:2105.14491, 2021

    Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks?arXiv preprint arXiv:2105.14491, 2021

    Pith/arXiv arXiv 2021

  40. [40]

    Sagess: Sampling graph denoising diffusion model for scalable graph generation.arXiv preprint arXiv:2306.16827, 2023

    Stratis Limnios, Praveen Selvaraj, Mihai Cucuringu, Carsten Maple, Gesine Reinert, and Andrew Elliott. Sagess: Sampling graph denoising diffusion model for scalable graph generation.arXiv preprint arXiv:2306.16827, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    Dimensionality reduction for visualizing single-cell data using umap.Nature biotechnology, 37(1):38–44, 2019

    Etienne Becht, Leland McInnes, John Healy, Charles-Antoine Dutertre, Immanuel WH Kwok, Lai Guan Ng, Florent Ginhoux, and Evan W Newell. Dimensionality reduction for visualizing single-cell data using umap.Nature biotechnology, 37(1):38–44, 2019

    2019

  42. [42]

    Data navigation on the encode portal.Nature Communications, 16(1):9592, 2025

    Meenakshi S Kagda, Bonita Lam, Casey Litton, Corinn Small, Cricket A Sloan, Emma Spragins, Forrest Tanaka, Ian Whaling, Idan Gabdank, Ingrid Youngworth, et al. Data navigation on the encode portal.Nature Communications, 16(1):9592, 2025

    2025

  43. [43]

    Chip-atlas 3.0: a data-mining suite to explore chromosome architecture together with large-scale regulome data.Nucleic Acids Research, 52 (W1):W45–W53, 2024

    Zhaonan Zou, Tazro Ohta, and Shinya Oki. Chip-atlas 3.0: a data-mining suite to explore chromosome architecture together with large-scale regulome data.Nucleic Acids Research, 52 (W1):W45–W53, 2024

    2024

  44. [44]

    Escape: database for integrating high-content published data collected from human and mouse embryonic stem cells.Database, 2013:bat045, 2013

    Huilei Xu, Caroline Baroukh, Ruth Dannenfelser, Edward Y Chen, Christopher M Tan, Yan Kou, Yujin E Kim, Ihor R Lemischka, and Avi Ma’ayan. Escape: database for integrating high-content published data collected from human and mouse embryonic stem cells.Database, 2013:bat045, 2013

    2013

  45. [45]

    Benchmark and integration of resources for the estimation of human transcription factor activities.Genome research, 29(8):1363–1375, 2019

    Luz Garcia-Alonso, Christian H Holland, Mahmoud M Ibrahim, Denes Turei, and Julio Saez- Rodriguez. Benchmark and integration of resources for the estimation of human transcription factor activities.Genome research, 29(8):1363–1375, 2019

    2019

  46. [46]

    Regnetwork: an integrated database of transcriptional and post-transcriptional regulatory networks in human and mouse.Database, 2015:bav095, 2015

    Zhi-Ping Liu, Canglin Wu, Hongyu Miao, and Hulin Wu. Regnetwork: an integrated database of transcriptional and post-transcriptional regulatory networks in human and mouse.Database, 2015:bav095, 2015

    2015

  47. [47]

    Trrust v2: an expanded reference database of human and mouse transcriptional regulatory interactions.Nucleic acids research, 46(D1):D380–D386, 2018

    Heonjong Han, Jae-Won Cho, Sangyoung Lee, Ayoung Yun, Hyojin Kim, Dasom Bae, Sunmo Yang, Chan Yeong Kim, Muyoung Lee, Eunbeen Kim, et al. Trrust v2: an expanded reference database of human and mouse transcriptional regulatory interactions.Nucleic acids research, 46(D1):D380–D386, 2018

    2018

  48. [48]

    Damian Szklarczyk, Rebecca Kirsch, Mikaela Koutrouli, Katerina Nastou, Farrokh Mehryary, Radja Hachilif, Annika L Gable, Tao Fang, Nadezhda T Doncheva, Sampo Pyysalo, et al. The string database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest.Nucleic acids research, 51(D1):D638–D646, 202...

    2023

  49. [49]

    hESChuman embryonic stem cells

  50. [50]

    hHEPhuman mature hepatocytes

  51. [51]

    mDCmouse dendritic cells

  52. [52]

    mESCmouse embryonic stem cells

  53. [53]

    mHSC-Emouse hematopoietic stem cellswith erythroid-lineage

  54. [54]

    mHSC-Lmouse hematopoietic stem cellswith lymphoid-lineage

  55. [55]

    Each cluster corresponds to a distinct cellular state, supporting the use of cell-level structure as the basis for discretization

    mHSC-GMmouse hematopoietic stem cellswith granulocyte-monocyte-lineage C0hESCC1C2C3 C0hHEPC1C2C3 C0mDCC1C2C3 C0mESCC1C2C3 C0mHSC-EC1C2C3 C0mHSC-GMC1C2C3 C0mHSC-LC1C2C3 Figure 4:Cell-state structure induced by clustering.Visualization of cell clusters obtained from scRNA-seq profiles using UMAP [41]. Each cluster corresponds to a distinct cellular state, s...