Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

TxPert: Leveraging Biochemical Relationships for Out-of-Distribution Transcriptomic Perturbation Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TxPert claims to predict unseen gene-perturbation effects at near-experimental fidelity by reading gene identity through biological knowledge graphs.

desk verdict Solid empirical paper with a load-bearing provenance gap: the headline SOTA depends on proprietary graphs that may encode the test labels. read the letter →

arxiv 2505.14919 v1 pith:SPIQXI2P submitted 2025-05-20 cs.LG q-bio.QM

Frederik Wenkel , Wilson Tu , Cassandra Masschelein , Hamed Shirzad , Cian Eastwood , Shawn T. Whitfield , Ihab Bendidi , Craig Russell
show 7 more authors
Liam Hodgson Yassir El Mesbahi Jiarui Ding Marta M. Fay Berton Earnshaw Emmanuel Noutahi Alisandra K. Denton
This is my paper
classification cs.LGq-bio.QM
keywords transcriptomicperturbationpredictionout-of-distributiongeneralizationgraphneuralnetworksknowledgegraphslatenttransferPerturb-seqsingle-celltranscriptomicsbatcheffect
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TxPert is a deep learning framework that predicts the full transcriptomic response to a genetic perturbation by combining a learned embedding of the unperturbed cell with a graph-neural-network embedding of the perturbed gene or genes. The paper claims it outperforms existing methods (GEARS, scLAMBDA) and a strong non-learned baseline on three out-of-distribution tasks: unseen single perturbations in a known cell line, unseen double perturbations, and single perturbations in an entirely unseen cell line. On unseen single perturbations, TxPert reaches or approaches the level of experimental reproducibility in three of the four cell lines tested (K562, Jurkat, HEPG2). This matters because exhaustive experimental screening is prohibitively expensive, so in silico predictions could prioritize which perturbations to test in drug discovery. The paper also argues for stricter evaluation: batch-matched controls and retrieval metrics are needed to avoid mistaking batch effects or generic stress responses for perturbation-specific predictive skill.

What carries the argument

The machinery is latent transfer over gene-gene knowledge graphs. A basal-state encoder (an MLP on control expression profiles, optionally with batch-matched control averaging) produces an embedding of the cell before perturbation; a GNN perturbation encoder maps each gene to an embedding by message passing over one or more interaction graphs, and the perturbation embedding is added to the basal embedding as a latent shift, with a decoder predicting the log-transformed expression profile. The multi-graph variant Exphormer-MG builds a union of edges across STRINGdb, GO, PxMap, and TxMap, using multi-hot edge-provenance features and an expander graph to support long-range attention. The paper identifies attention-based GNNs as suited for noisy graphs and multi-hop message passing as compensating for edge dropout; four-layer models stay stable when up to 60% of edges are removed.

What would settle it

Check whether the genes, cell lines, and perturbation identities used to construct PxMap and TxMap appear in the held-out splits of the Replogle, Norman, and Nadig datasets; if they do, retrain TxPert without those graphs and compare. A second direct test is to rerun the full TxPert pipeline using only STRINGdb and GO on public data and see whether it still beats the general baseline and approaches experimental reproducibility.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single framework, TxPert, can generalize to perturbations it has never seen by reading the perturbed gene's identity through multiple biological knowledge graphs. For unseen single perturbations within a cell line, TxPert reports state-of-the-art Pearson $\Delta$ correlation, beating GEARS, scLAMBDA, and a non-learned general baseline, and matching or approaching experimental reproducibility in K562, Jurkat, and HEPG2 cells. For double perturbations, it outperforms the additive baseline and both published methods. For transfer to an unseen cell line, TxPert beats both the general baseline and the adapted scLAMBDA baseline in all four held-out lines. The authors treat the graph priors as essential: progressively rewiring the STRINGdb graph degrades performance, and combining STRINGdb, GO, PxMap, and TxMap into a multi-graph Exphormer transformer improves performance beyond any single graph.

Load-bearing premise

The load-bearing premise is that the proprietary PxMap and TxMap graphs were built from perturbation screens that do not overlap the held-out test perturbations or cell lines; if those graphs already encode the response to a test perturbation, the reported 'unseen' generalization is inflated.

Editorial extensions

If this is right

  • If the results hold, unseen single perturbations can be prioritized in silico at a fidelity approaching the replicate-to-replicate consistency of the assay itself in well-characterized cell lines.
  • Double-perturbation response prediction from seen singles can replace costly combinatorial screens for first-pass prioritization of candidate gene pairs.
  • Zero-shot prediction to an unseen cell line is feasible for seen perturbations, making cross-context virtual screening a realistic target as cell-line diversity in training data grows.
  • Evaluation standards using batch-matched controls and retrieval metrics should be adopted by later perturbation models, since global mean baselines capture generic stress responses rather than perturbation-specific signal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the paper does not show that the proprietary PxMap and TxMap graphs were built without overlap with the held-out perturbations, a public-data-only variant of TxPert is the natural test of whether the out-of-distribution claim reflects genuine generalization or partly information leakage.
  • Editorial inference: the strength of the mean baseline implies that correlation-style metrics can look good even when a model is merely reproducing generic stress biology, so retrieval metrics and per-perturbation specificity should carry more weight in future evaluations.
  • Editorial inference: a testable extension is to feed larger, multi-cell-line perturbation atlases into the basal-state encoder; the paper's own unseen-cell-line results suggest this would directly reduce the gap to experimental reproducibility.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. TxPert is a deep learning framework for predicting transcriptomic responses to genetic perturbations in three out-of-distribution settings: unseen single-gene perturbations within a known cell line, unseen double perturbations, and perturbations transferred to unseen cell lines. The model combines a basal-state encoder with a graph neural network perturbation encoder operating on multiple gene-gene interaction graphs, including curated sources (STRINGdb, GO) and two proprietary graphs derived from perturbation screens (PxMap, TxMap). The authors report that TxPert outperforms GEARS, scLAMBDA, and a non-learned general baseline in these settings, and that it approaches or matches experimental reproducibility for unseen single perturbations in some cell lines. The paper also proposes evaluation practices: batch-matched controls, stratified knowledge-level analysis, retrieval metrics, and a split-half experimental reproducibility estimate.

Significance. If the central claims hold, this is a useful contribution to perturbation prediction: the evaluation framework is more careful than much of the prior literature, with batch-matched controls, retrieval metrics, an explicit non-learned baseline, ablation studies, and an estimate of experimental reproducibility. The authors ship code for the public-data parts of the pipeline and are honest about limitations (e.g., the failure to predict the perturbation target's own downregulation). The proprietary-graph provenance question, the absence of error bars, and the missing linear baseline are the main factors preventing acceptance of the state-of-the-art claims as stated.

major comments (3)
  1. [Sections 2.2, 4.3.2, 5] The central OOD claim for unseen single perturbations (Fig. 3A) rests on TxMap, a graph built from proprietary single-cell transcriptomics perturbation screens (Section 2.2), and Fig. 4D shows that adding TxMap improves performance. Section 4.3.2 states that screen-derived graphs are computed as pairwise cosine similarities of gene embeddings derived from perturbation screens, but the manuscript does not report whether the internal screens used to build TxMap (or PxMap) include the genes, perturbations, or cell lines that appear in the Replogle, Nadig, or Norman evaluation splits. If any internal screen covered an ostensibly held-out perturbation, the graph edge weights encode the observed transcriptional response of that perturbation, and a GNN can retrieve it via neighboring nodes, inflating the reported OOD performance. Because TxMap is proprietary and not released (Section 5), this cannot be verified externally. The authors should provide a provenance audit: a list of the genes, perturbation types, and cell lines covered by the internal screens, and an overlap analysis against the evaluation splits; alternatively, they should report the main results with TxMap and PxMap excluded and show that the SOTA conclusion survives.
  2. [Sections 2.2, 4.1.3, 4.2.1] The paper tunes each architecture individually and reports the best model variant per task (Section 2.2: 'we tuned each architecture individually and report the best model variant per task'; Section 4.1.3), and the main results in Figures 3A, 3B, and 3C are point estimates with no error bars or confidence intervals. This makes it impossible to assess whether the reported differences between TxPert and baselines, or between TxPert and the experimental reproducibility estimate, are statistically meaningful; the claim in Section 2.3 that TxPert 'is competitive with experimental reproducibility' is a comparison of point estimates. The authors should report means and confidence intervals across multiple seeds, and ideally a table of all model variants per task, to address the risk of selection overfitting.
  3. [Sections 4.2.3 and 3] No linear baseline is included, despite the paper's own introduction noting that independent benchmarks have found linear methods to be strong (Ahlmann-Eltze et al., 2024; Kernfeld et al., 2023). The General Baseline is a non-learned mean/additive model, not a fitted linear model, and a linear regression baseline (e.g., per-gene ridge regression with perturbation one-hot features) could be substantially stronger. To support the 'state-of-the-art' claim, the evaluation framework should include a simple learned linear baseline trained under the same data splits.
minor comments (5)
  1. [Section 4.2.2] In the metric definitions, 'Person ∆' is a typo and should read 'Pearson ∆'.
  2. [Section 4.3.2] In the cosine similarity formula, the denominator repeats ∥xgi ∥; the second norm should be ∥xgj ∥.
  3. [Section 5] The sentence 'All data used here is already publicly available; with the exception of the PxMap and TxMap graphs' is self-contradictory; rephrase to clarify that all evaluation data are publicly available and only the two graphs are proprietary.
  4. [Figure 1D and Section 4.3.3] The legend states that expected random performance is 0.9 for the normalized retrieval metric because the 0.9 quantile is plotted; this dependence should be stated explicitly in the main text to avoid confusion.
  5. [Appendix A, 'Graph Attention Network'] The attention formula introduces the notation ATTθ,W(hu,hv) but then defines auv and α(q,k); the notation should be made consistent.

Circularity Check

0 steps flagged · score 2.0 of 10

No demonstrated circularity; TxPert is trained end-to-end on held-out perturbations, though the proprietary TxMap/PxMap graphs pose an unquantified data-independence risk.

full rationale

Walking the derivation chain in Sections 4.1–4.2, the prediction is produced by a learned decoder g_phi applied to a latent shift of a basal-state embedding and GNN-derived perturbation embeddings; all parameters are optimized with MSE on training perturbations, and test perturbations are held out by perturbation ID (Section 4.2.1), so no output is a renamed fitted constant. The General Baseline and experimental-reproducibility estimates are explicit comparators, not model outputs. The SOTA claim is supported by held-out evaluations (Fig. 3) and ablations (Fig. 4), not by citation alone. The main caveat is that TxMap and PxMap are proprietary graphs built from perturbation screens (Sections 2.2 and 4.3.2) and are not released (Section 5). If the internal screens used to construct TxMap overlap the Replogle, Nadig, or Norman test perturbations or cell lines, the graph edges could encode the very transcriptional responses the model is asked to predict, which would inflate the OOD results; however, the paper never asserts such an overlap, and the manuscript's equations do not exhibit a by-construction identity between graph weights and test labels. The absence of a provenance audit is a data-independence risk for external verification rather than a demonstrated circular step, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim depends on two proprietary knowledge graphs that are not publicly accessible, several empirically chosen preprocessing thresholds, and per-task model selection. These are the main elements that a reader cannot independently audit or reproduce.

free parameters (4)
  • Edge filtering threshold for screen-based graphs (top 1%) = 1% (top edges by absolute weight)
    Section 4.3.2: 'Filtering configurations were optimized empirically. We found that the most performant configuration involved selecting for the top 1% of edges by (absolute) weight for screen-based graphs.' This hand-chosen threshold affects the graph input and thus the model's predictions.
  • Maximum in-degree for curated graph filtering (20 incoming nodes) = 20 incoming nodes per target
    Section 4.3.2: 'For all other graph types we (additionally) filtered for no more than 20 incoming nodes by target.' Empirically chosen.
  • TxPert model variant per task (Exphormer-MG, GAT-MultiLayer, GAT-Hybrid, etc.) = Varies by task (best on validation)
    Section 2.2 and 4.1.3: 'we tuned each architecture individually and report the best model variant per task in the main text.' This model selection uses validation Pearson ∆.
  • Basal state matching and averaging settings = On/off chosen per task; averaging used in standard config
    Section 4.1.1 and Figure 7: basal matching and averaging consistently improved performance and are part of standard configuration, but the on/off choice for unseen cell lines was empirical.
assumptions (4)
  • domain assumption Gene interaction knowledge graphs (STRINGdb, GO, PxMap, TxMap) encode relationships that provide useful inductive bias for predicting transcriptional perturbation responses.
    The entire TxPert design (Section 4.1.2) rests on this. Ablations (Fig 4A,B) show the graph structure matters, but the biological validity of the proprietary graphs is assumed.
  • ad hoc to paper TxMap and PxMap are constructed from perturbation screens without overlapping the evaluation perturbations, so the OOD test remains clean.
    Section 4.3.2 describes these graphs but does not report whether the internal screens share genes, cell lines, or perturbation data with the public test sets (Replogle, Norman, Nadig). This is load-bearing for the 'unseen perturbation' claim.
  • domain assumption Batch-matched controls (matching and averaging) provide a faithful basal state that removes batch effects without removing perturbation signal.
    Section 2.1 and 4.1.1 justify this with control correlation analyses, but the correction is assumed to be sufficient for the reported Pearson ∆ comparisons.
  • standard math Standard definitions of GNNs, attention, and MSE loss are correct and applicable.
    Model definitions in Appendix A are standard; no new mathematics.
invented entities (2)
  • PxMap
    purpose: Proprietary Recursion gene-gene graph derived from microscopy phenomics, used as a perturbation prior in the best-performing Exphormer-MG configuration.
    Not publicly available (Section 5); the paper provides no external access or independent validation of its construction. Central SOTA results depend on it.
  • TxMap
    purpose: Proprietary Recursion gene-gene graph derived from single-cell transcriptomics perturbation screens, used as a perturbation prior; also used in GAT-Hybrid and GAT-MultiLayer variants.
    Not publicly available; because it is derived from transcriptomic perturbation screens, it may encode information overlapping the prediction target, and the paper does not demonstrate otherwise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TxPert: Leveraging Biochemical Relationships for Out-of-Distribution Transcriptomic Perturbation Prediction." pith.science (2026). https://pith.science/paper/SPIQXI2P

@misc{pith2026250514919,
  author       = {Pith},
  title        = {Pith review of: TxPert: Leveraging Biochemical Relationships for Out-of-Distribution Transcriptomic Perturbation Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPIQXI2P}},
  note         = {Machine review of arXiv:2505.14919}
}
read the original abstract

Accurately predicting cellular responses to genetic perturbations is essential for understanding disease mechanisms and designing effective therapies. Yet exhaustively exploring the space of possible perturbations (e.g., multi-gene perturbations or across tissues and cell types) is prohibitively expensive, motivating methods that can generalize to unseen conditions. In this work, we explore how knowledge graphs of gene-gene relationships can improve out-of-distribution (OOD) prediction across three challenging settings: unseen single perturbations; unseen double perturbations; and unseen cell lines. In particular, we present: (i) TxPert, a new state-of-the-art method that leverages multiple biological knowledge networks to predict transcriptional responses under OOD scenarios; (ii) an in-depth analysis demonstrating the impact of graphs, model architecture, and data on performance; and (iii) an expanded benchmarking framework that strengthens evaluation standards for perturbation modeling.

Figures

Figures reproduced from arXiv: 2505.14919 by the authors.

Figure 1
Figure 1. A) Pearson correlation of aggregated control gene expression profiles within and across [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. A) The TxPert architecture predicts post-perturbation gene expression by combining two modules: (1) a basal state encoder that maps batch-matched control profiles into a latent embedding, and (2) a Graph Neural Network (GNN) that learns perturbation embeddings from a gene-gene interaction graph. Perturbation embeddings are applied to the basal embedding, and the resulting latent representation is decoded to produce … view at source ↗
Figure 3
Figure 3. A) Performance of TxPert compared to GEARS and scLAMBDA on predicting unseen single perturbations within a known cell type. Horizontal bars indicate general baseline, a batch-informed model (capturing potential confounding), and experimental reproducibility (see Section 4.1). B) Comparison of TxPert, GEARS and scLAMBDA in predicting double perturbation effects from known singles. C) Comparison of TxPert and scLAMBDA… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Ablation studies for unseen perturbation effect prediction on K562. A) Performance of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Investigation into strengths and weaknesses of our models. A) Breakdown of Pearson ∆ by [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: A) Visualization of TxPert performance in modeling the knockdown of an example held out test gene, TSR2 on primary and secondary neighbors (up to 30 total, prioritized by weight). Neighbors and edges are from the STRINGdb graph used for modeling. Colors indicate the lo…
Figure 7
Figure 7. Figure 7: Ablation study of modeling choices for unseen cell line OOD task on K562. We compare [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Normalized retrieval between true perturbant replicates in different biological contexts. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: Performance of TxPert compared to GEARS and scLAMBDA on retrieving unseen single perturbations using predicted perturbation effect within a known cell type. Horizontal bars indicate general baseline, a batch-informed model (capturing potential confounding), and experim…
Figure 10
Figure 10. Figure 10: The performance of scLAMBDA and TxPert at the task of single perturbations in unseen cell lines. Horizontal bars put the predictions in the context of the general baseline, a learned model making predictions on the basis of batch information (in case of confounding be…
Figure 11
Figure 11. Figure 11: The performance of GEARS, scLAMBDA and TxPert for prediction of double perturbations on Norman for various OOD settings. We report overall Pearson ∆ and performance when both, one or none of the individual perturbations that constitute the double have been seen during…
Figure 12
Figure 12. Figure 12: Perturbation effect prediction of unseen perturbations in K562: A) Randomly downsampling [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Perturbation effect prediction of unseen perturbations in K562: This plot compares a 4-layer [PITH_FULL_IMAGE:figures/full_fig_p028_13.png]
Figure 14
Figure 14. Figure 14: Exploration into potential relationships between perturbation-target metadata and the [PITH_FULL_IMAGE:figures/full_fig_p029_14.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. PertReason: A Knowledge-Grounded Benchmark and Framework for Cell-State-Conditioned Mechanistic Reasoning of Perturbation Effects

    cs.LG 2026-07 conditional novelty 6.0 of 10

    PertReasonQA scores AI models on cell-state-conditioned mechanistic reasoning about perturbation effects, and PertReasonLM, trained with reasoning supervision, reaches 0.736 balanced accuracy and 0.976 edge recall ver...

  2. Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction

    cs.CE 2026-02 conditional novelty 6.0 of 10

    AdaPert improves transcriptional perturbation prediction by conditioning sparse knowledge-graph subgraphs on each perturbed gene and training with signal/noise separation losses, achieving the best DEG recovery on K56...

Reference graph

Works this paper leans on

20 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Deep learning-based predictions of gene perturbation effects do not yet outperform simple linear methods.BioRxiv, pp

    Ahlmann-Eltze, C., Huber, W., and Anders, S. Deep learning-based predictions of gene perturbation effects do not yet outperform simple linear methods.BioRxiv, pp. 2024–09,

    work page 2024

  2. [6]

    M., Zhou, Y., Crepaldi, L., Usluer, S., Dunham, A., Braunger, J

    Feng, C., Peets, E. M., Zhou, Y., Crepaldi, L., Usluer, S., Dunham, A., Braunger, J. M., Su, J., Strauss, M. E., Muraro, D., et al. A genome-scale single cell crispri map of trans gene regulation across human pluripotent stem cell lines.bioRxiv, pp. 2024–11,

    work page 2024

  3. [9]

    K., Mathias, S

    Sheils, T. K., Mathias, S. L., Kelleher, K. J., Siramshetty, V. B., Nguyen, D.-T., Bologa, C. G., Jensen, L. J., Vidovi´ c, D., Koleti, A., Sch¨ urer, S. C., et al. Tcrd and pharos 2021: mining the human proteome for disease biology.Nucleic Acids Research, 49(D1):D1334–D1346,

    work page 2021

  4. [10]

    A theory for compressibility of graph transformers for transductive learning.arXiv preprint arXiv:2411.13028, 2024a

    Shirzad, H., Lin, H., Velingker, A., Venkatachalam, B., Woodruff, D., and Sutherland, D. A theory for compressibility of graph transformers for transductive learning.arXiv preprint arXiv:2411.13028, 2024a. Shirzad, H., Lin, H., Venkatachalam, B., Velingker, A., Woodruff, D. P., and Sutherland, D. J. Even sparser graph transformers.Advances in Neural Infor...

    arXiv

  5. [12]

    Cellplm: pre-training of cell language model beyond single cells.BioRxiv, pp

    Wen, H., Tang, W., Dai, X., Ding, J., Jin, W., Xie, Y., and Tang, J. Cellplm: pre-training of cell language model beyond single cells.BioRxiv, pp. 2023–10,

    work page 2023

  6. [13]

    Overcoming Oversmoothness in Graph Convolutional Networks via Hybrid Scattering Networks

    Wenkel, F., Min, Y., Hirn, M., Perlmutter, M., and Wolf, G. Overcoming oversmoothness in graph convolutional networks via hybrid scattering networks.arXiv preprint arXiv:2201.08932,

    work page Pith review arXiv

  7. [14]

    Cayley graph propagation.arXiv preprint arXiv:2410.03424,

    Wilson, J., Bechler-Speicher, M., and Veliˇ ckovi´ c, P. Cayley graph propagation.arXiv preprint arXiv:2410.03424,

    arXiv

  8. [15]

    R., Hill, A., and Moccia, R

    Wong, D. R., Hill, A., and Moccia, R. Simple controls exceed best deep learning algorithms and reveal foundation model effectiveness for predicting genetic perturbations.bioRxiv, pp. 2025–01,

    work page 2025

Show all 20 references
  1. [18]

    A., Borja, R

    Zhang, J., Ubas, A. A., Borja, R. de, Svensson, V., Thomas, N., Thakar, N., Lai, I., Winters, A., Khan, U., Jones, M. G., et al. Tahoe-100m: A giga-scale single-cell perturbation atlas for context-dependent gene function and cellular modeling.bioRxiv, pp. 2025–02,

    2025

  2. [19]

    , vN ⋆ } and edges E⊂V×V representing gene-gene interactions

    21 TxPert Appendix : Leveraging Biochemical Relationships for Out-of-Distribution Transcriptomic Perturbation Prediction A GNN perturbation encoder model details General graph network structure.We consider a simple weighted graph G = (V, E, w) with a set of N ⋆ ∈N nodes (or ge...

    2021

  3. [20]

    Exphormer

    =f L ◦f L−1 ◦ · · · ◦f1(H0). This model uses an attention mechanism to aggregate information during message-passing (MP). For every edge, we learn an attention score that characterizes its importance that is then normalized at every node across incoming edges. On the node leve...

    2024

  4. [2004]

    Z., and Szalai, B

    Csendes, G., Szalay, K. Z., and Szalai, B. Benchmarking a foundational cell model for post-perturbation rnaseq prediction.bioRxiv, pp. 2024–09,

    2024

  5. [2016]

    Dwivedi, V. P. and Bresson, X. A generalization of transformer networks to graphs.arXiv preprint arXiv:2012.09699,

    2012 arXiv

  6. [2018]

    Modeling and predicting single-cell multi-gene perturbation responses with sclambda.bioRxiv, pp

    Wang, G., Liu, T., Zhao, J., Cheng, Y., and Zhao, H. Modeling and predicting single-cell multi-gene perturbation responses with sclambda.bioRxiv, pp. 2024–12,

    2024

  7. [2019]

    M., Nassar, M., Osi´ nski, B., Eksi, R., Zhang, K., and Graepel, T

    Wu, Y., Wershof, E., Schmon, S. M., Nassar, M., Osi´ nski, B., Eksi, R., Zhang, K., and Graepel, T. Perturbench: Benchmarking machine learning models for cellular perturbation analysis.arXiv preprint arXiv:2408.10609,

  8. [2021]

    Link prediction on multilayer networks through learning of within-layer and across-layer node-pair structural features and node embedding similarity

    Zangari, L., Mandaglio, D., and Tagarelli, A. Link prediction on multilayer networks through learning of within-layer and across-layer node-pair structural features and node embedding similarity. In Proceedings of the ACM Web Conference 2024, pp. 924–935,

    2024

  9. [2022]

    K., Bronstein, M

    Di Giovanni, F., Rusch, T. K., Bronstein, M. M., Deac, A., Lackenby, M., Mishra, S., and Veliˇ ckovi´ c, P. How does over-squashing affect the power of gnns?arXiv preprint arXiv:2306.03589,

    arXiv

  10. [2023]

    The reactome pathway knowledgebase 2024.Nucleic acids research, 52(D1): D672–D678,

    Milacic, M., Beavers, D., Conley, P., Gong, C., Gillespie, M., Griss, J., Haw, R., Jassal, B., Matthews, L., May, B., et al. The reactome pathway knowledgebase 2024.Nucleic acids research, 52(D1): D672–D678,

    2024

  11. [2024]

    B., Mesbahi, Y

    Bendidi, I., Whitfield, S., Kenyon-Dean, K., Yedder, H. B., Mesbahi, Y. E., Noutahi, E., and Denton, A. K. Benchmarking transcriptomics foundation models for perturbation analysis: one pca still rules them all.arXiv preprint arXiv:2410.13956,

    arXiv

  12. [2025]

    S., Battle, A., and Cahan, P

    Kernfeld, E., Yang, Y., Weinstock, J. S., Battle, A., and Cahan, P. A systematic comparison of computational methods for expression forecasting.BioRxiv, pp. 2023–07,

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.