pith. sign in

arxiv: 2606.24940 · v1 · pith:IDOX7WVEnew · submitted 2026-06-22 · 🧬 q-bio.GN · cs.AI· cs.LG

Stable-Shift: Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations

Sajib Acharjee Dip , Liqing Zhang This is my paper

Pith reviewed 2026-06-26 05:53 UTC · model grok-4.3

classification 🧬 q-bio.GN cs.AIcs.LG
keywords transcriptional response predictiongene perturbationunseen geneslow-rank basisbiological contextgraph convolutionPerturb-seqfunctional genomics
0
0 comments X

The pith

Stable-Shift predicts transcriptional responses to gene perturbations never seen during training by mapping biological context into a low-rank response basis.

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

The paper presents a method that aggregates single-cell perturbation data into expression shifts and learns a low-rank basis from only the training genes. It then uses graph convolution on features from protein interactions, network properties, expression statistics, and gene ontology to place an unseen gene into that basis and predict its response. This approach is evaluated on the K562 Perturb-seq dataset where it shows improved cosine similarity over a prior method. A sympathetic reader would care because it offers a way to generalize predictions beyond the set of experimentally perturbed genes.

Core claim

Stable-Shift aggregates single-cell measurements into perturbation-level expression shifts, fits a low-rank response basis using training perturbations only, and predicts an unseen gene's coordinates in that basis from biological context. The context combines STRING interactions, network structure, control-cell expression statistics, and Gene Ontology annotations; the evaluated implementation uses graph convolution to integrate these inputs.

What carries the argument

Low-rank response basis fitted from training perturbations, with coordinates for unseen genes predicted by graph convolution over biological context features.

If this is right

  • Improved cosine similarity of 0.592 compared to 0.569 for GEARS on the K562 benchmark.
  • Higher Spearman correlation and top-gene precision among evaluated methods.
  • Consistent ordering across graph-aware, residualized, gene-space, and Norman-dataset comparisons.
  • Support for further study of biologically structured latent-response prediction.

Where Pith is reading between the lines

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

  • Such methods could extend to other cell types if the biological context features generalize.
  • Combining with more advanced graph models might improve performance in sparse graph neighborhoods.
  • Testing on larger perturbation libraries would reveal scalability limits.
  • The low gene-space accuracy noted suggests the method works better in latent space than direct expression prediction.

Load-bearing premise

Biological context features drawn from STRING interactions, network structure, control-cell expression statistics, and Gene Ontology annotations are sufficiently informative to accurately locate an unseen gene inside the low-rank response basis that was fitted exclusively from training perturbations.

What would settle it

Observing that on a held-out set of genes Stable-Shift cosine similarity drops below the GEARS baseline would indicate the context features do not reliably locate genes in the basis.

Figures

Figures reproduced from arXiv: 2606.24940 by Liqing Zhang, Sajib Acharjee Dip.

Figure 1
Figure 1. Figure 1: Stable-Shift workflow. Perturbation-level expression shifts define a training-only response basis, while complementary [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative analysis of Stable-Shift on unseen perturbations. Panels (a–c) compare observed and predicted shifts [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Predicting transcriptional responses to genetic perturbations could reduce the experimental burden of functional genomics, but extrapolation to genes that were never perturbed during training remains difficult. We present Stable-Shift, a structured method for estimating unseen-gene responses. Stable-Shift aggregates single-cell measurements into perturbation-level expression shifts, fits a low-rank response basis using training perturbations only, and predicts an unseen gene's coordinates in that basis from biological context. The context combines STRING interactions, network structure, control-cell expression statistics, and Gene Ontology annotations; the evaluated implementation uses graph convolution to integrate these inputs. On the supplied K562 Perturb-seq benchmark, Stable-Shift obtained 0.592 cosine similarity, compared with 0.569 for GEARS, together with higher Spearman correlation and top-gene precision among the evaluated methods. Its mean cosine similarity over five unseen-gene splits was 0.589 +/- 0.008. The same ordering was observed in the supplied graph-aware, residualized, gene-space, and Norman-dataset comparisons. These results support further study of biologically structured latent-response prediction, while the lower gene-space accuracy and sensitivity to sparse graph neighborhoods limit the scope of the present conclusions.

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 manuscript presents Stable-Shift, a method for predicting transcriptional responses to unseen gene perturbations. It aggregates single-cell measurements into perturbation-level shifts, fits a low-rank response basis using only training perturbations, and predicts an unseen gene's coordinates in this basis using graph convolution over biological context features including STRING interactions, network structure, control-cell expression, and Gene Ontology annotations. On the K562 Perturb-seq benchmark, it achieves a cosine similarity of 0.592 compared to 0.569 for GEARS, with similar improvements in other metrics and consistent ordering across multiple evaluation variants and datasets.

Significance. If the results hold, the approach offers a promising way to extrapolate to unperturbed genes by leveraging external biological knowledge in a structured latent space, which could substantially reduce the experimental costs in functional genomics. The provision of standard deviations across five splits and consistent performance across variants strengthens the empirical support. The low circularity (response basis independent of test data) is a positive aspect.

major comments (2)
  1. [Abstract] Abstract: The central performance claim depends on the biological context features being sufficiently informative to accurately regress coordinates in the low-rank basis fitted from training data. However, no ablation of individual feature sources (STRING, GO, etc.) or degree-stratified metrics for sparse neighborhoods are provided, despite the abstract noting sensitivity to sparse graph neighborhoods. This leaves open whether the modest improvement (0.592 vs 0.569 cosine similarity) is attributable to the structured prediction or other factors.
  2. [Methods] Methods: Details on the hyperparameter selection procedure, data exclusion criteria, and any statistical tests for the reported metrics (e.g., mean cosine similarity 0.589 +/- 0.008 over five splits) are absent, which is necessary to fully substantiate the soundness of the performance claims.
minor comments (1)
  1. [Abstract] The abstract could more explicitly state the rank of the response basis and the specific graph convolution architecture parameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. The comments highlight opportunities to strengthen the empirical support and methodological transparency of Stable-Shift. We address each major comment below and commit to revisions that directly incorporate the requested analyses and details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claim depends on the biological context features being sufficiently informative to accurately regress coordinates in the low-rank basis fitted from training data. However, no ablation of individual feature sources (STRING, GO, etc.) or degree-stratified metrics for sparse neighborhoods are provided, despite the abstract noting sensitivity to sparse graph neighborhoods. This leaves open whether the modest improvement (0.592 vs 0.569 cosine similarity) is attributable to the structured prediction or other factors.

    Authors: We agree that the absence of feature ablations and degree-stratified metrics leaves the source of the observed gains incompletely substantiated. Although the abstract already flags sensitivity to sparse neighborhoods and the main text reports consistent ordering across multiple evaluation variants, these do not replace explicit ablations. In the revised manuscript we will add (i) an ablation table removing each context source (STRING, GO, expression statistics, network structure) in turn and (ii) performance metrics stratified by graph degree (e.g., low-, medium-, and high-degree bins). These additions will allow readers to assess whether the modest but consistent improvement is driven by the biologically structured prediction. revision: yes

  2. Referee: [Methods] Methods: Details on the hyperparameter selection procedure, data exclusion criteria, and any statistical tests for the reported metrics (e.g., mean cosine similarity 0.589 +/- 0.008 over five splits) are absent, which is necessary to fully substantiate the soundness of the performance claims.

    Authors: We acknowledge that the current Methods section omits these procedural details. The reported mean and standard deviation are computed across five independent unseen-gene splits, yet the hyperparameter search strategy, exclusion rules, and any formal statistical comparisons are not described. In the revision we will insert a new subsection that specifies (i) the hyperparameter selection procedure and validation approach, (ii) all data exclusion criteria applied during preprocessing and split construction, and (iii) the exact statistical procedures used to obtain means, standard deviations, and any significance assessments. This will enable full reproducibility and evaluation of the performance claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses independent external features

full rationale

The paper fits a low-rank response basis exclusively from training perturbations and regresses unseen-gene coordinates in that basis from external biological context features (STRING interactions, GO annotations, network structure, control expression). These features are not functions of the response measurements themselves, so the prediction step does not reduce to a fit or self-definition by construction. No self-citation chains, ansatz smuggling, or renaming of known results appear in the load-bearing steps. The setup is a standard supervised extrapolation task whose validity hinges on feature informativeness rather than tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard machine-learning assumptions for genomics plus two domain assumptions about low-rank structure and context-response correlation; no new entities are postulated and the listed free parameters are typical modeling choices whose exact values are not reported in the abstract.

free parameters (2)
  • response basis rank
    Dimension of the low-rank basis fitted to training perturbation shifts; must be chosen or tuned and directly affects prediction capacity.
  • graph convolution architecture parameters
    Layer widths, aggregation functions, and training hyperparameters of the network that maps biological context to basis coordinates.
axioms (2)
  • domain assumption Transcriptional responses to perturbations admit a useful low-rank approximation.
    Invoked when the method collapses single-cell data into perturbation-level shifts and fits the basis used for all subsequent predictions.
  • domain assumption Biological context features (STRING, network structure, expression statistics, GO) are correlated with a gene's perturbation response profile.
    This premise justifies using the context features to predict the unseen gene's coordinates in the learned basis.

pith-pipeline@v0.9.1-grok · 5746 in / 1618 out tokens · 40990 ms · 2026-06-26T05:53:27.907574+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

37 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Abrar Rahman Abir, Sajib Acharjee Dip, and Liqing Zhang. 2025. CFM-GP: Unified Conditional Flow Matching to Learn Gene Perturbation Across Cell Types.arXiv preprint arXiv:2508.08312(2025)

    work page arXiv 2025

  2. [2]

    Britt Adamson, Thomas M Norman, Marco Jost, Min Y Cho, James K Nuñez, Yuwen Chen, Jacqueline E Villalta, Luke A Gilbert, Max A Horlbeck, Marco Y Hein, et al. 2016. A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response.Cell167, 7 (2016), 1867– 1882

    2016

  3. [3]

    Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. 2000. Gene ontology: tool for the unification of biology.Nature genetics25, 1 (2000), 25–29

    2000

  4. [4]

    Zhenyu Bi, Sajib Acharjee Dip, Daniel Hajialigol, Sindhura Kommu, Hanwen Liu, Meng Lu, and Xuan Wang. 2024. Ai for biomedicine in the era of large language models.arXiv preprint arXiv:2403.15673(2024)

    work page arXiv 2024

  5. [5]

    Hal Caswell. 2000. Prospective and retrospective perturbation analyses: their roles in conservation biology.Ecology81, 3 (2000), 619–627

    2000

  6. [6]

    Yifei Chen, Yi Li, Rajiv Narayan, Aravind Subramanian, and Xiaohui Xie. 2016. Gene expression inference with deep learning.Bioinformatics32, 12 (2016), 1832–1839

    2016

  7. [7]

    Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul, Uddip Acharjee Shuvo, Muhit Islam Emon, Xuan Wang, and Liqing Zhang. 2025. LLM4Cell: A Survey of Large Language and Agentic Models for Single-Cell Biology.arXiv preprint arXiv:2510.07793(2025)

    work page arXiv 2025

  8. [8]

    Atray Dixit, Oren Parnas, Biyu Li, Jenny Chen, Charles P Fulco, Livnat Jerby- Arnon, Nemanja D Marjanovic, Danielle Dionne, Tyler Burks, Raktima Ray- chowdhury, et al. 2016. Perturb-Seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens.cell167, 7 (2016), 1853–1866

    2016

  9. [9]

    Mirela Domijan, Paul E Brown, Boris V Shulgin, and David A Rand. 2016. PeTTSy: a computational tool for perturbation analysis of complex systems biology models. BMC bioinformatics17, 1 (2016), 124

    2016

  10. [10]

    Mohammed Eslami, Amin Espah Borujeni, Hamed Eramian, Mark Weston, George Zheng, Joshua Urrutia, Carolyn Corbet, Diveena Becker, Paul Maschhoff, Katie Clowers, et al. 2022. Prediction of whole-cell transcriptional response with machine learning.Bioinformatics38, 2 (2022), 404–409

    2022

  11. [11]

    Christopher Heje Grønbech, Maximillian Fornitz Vording, Pascal N Timshel, Casper Kaae Sønderby, Tune H Pers, and Ole Winther. 2020. scVAE: variational auto-encoders for single-cell gene expression data.Bioinformatics36, 16 (2020), 4415–4422

    2020

  12. [12]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. InProceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, New York, NY, USA, 855–864

    2016

  13. [13]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017), 1024–1034

    2017

  14. [14]

    Yuge Ji, Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. 2021. Machine learning for perturbational single-cell omics.Cell Systems12, 6 (2021), 522–537

    2021

  15. [15]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  16. [16]

    Peter Kohl, Edmund J Crampin, TA Quinn, and Denis Noble. 2010. Systems biology: an approach.Clinical Pharmacology & Therapeutics88, 1 (2010), 25–33

    2010

  17. [17]

    Mohammad Lotfollahi, Anna Klimovskaia Susmelj, Carlo De Donno, Leon Hetzel, Yuge Ji, Ignacio L Ibarra, Sanjay R Srivatsan, Mohsen Naghipourfar, Riza M Daza, Beth Martin, et al. 2023. Predicting cellular responses to complex perturbations in high-throughput screens.Molecular systems biology19, 6 (2023), MSB202211517

    2023

  18. [18]

    Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. 2019. scGen predicts single-cell perturbation responses.Nature methods16, 8 (2019), 715–721

    2019

  19. [19]

    Qin Ma and Dong Xu. 2022. Deep learning shapes single-cell data analysis.Nature reviews Molecular cell biology23, 5 (2022), 303–304

    2022

  20. [20]

    Kartik M Mani, Celine Lefebvre, Kai Wang, Wei Keat Lim, Katia Basso, Riccardo Dalla-Favera, and Andrea Califano. 2008. A systems biology approach to pre- diction of oncogenes and molecular perturbation targets in B-cell lymphomas. Molecular systems biology4 (2008), 169

    2008

  21. [21]

    Evan J Molinelli, Anil Korkut, Weiqing Wang, Martin L Miller, Nicholas P Gauthier, Xiaohong Jing, Poorvi Kaushik, Qin He, Gordon Mills, David B Solit, et al. 2013. Perturbation biology: inferring signaling networks in cellular systems.PLoS computational biology9, 12 (2013), e1003290

    2013

  22. [22]

    Joseph M Replogle, Reuben A Saunders, Angela N Pogson, Jeffrey A Hussmann, Alexander Lenail, Alina Guna, Lauren Mascibroda, Eric J Wagner, Karen Adelman, Gila Lithwick-Yanai, et al. 2022. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq.Cell185, 14 (2022), 2559–2575

    2022

  23. [23]

    Yusuf Roohani, Kexin Huang, and Jure Leskovec. 2024. Predicting transcriptional outcomes of novel multigene perturbations with GEARS.Nature Biotechnology 42, 6 (2024), 927–935

    2024

  24. [24]

    Heba Z Sailem, Jens Rittscher, and Lucas Pelkmans. 2020. KCML: a machine- learning framework for inference of multi-scale gene functions from genetic perturbation screens.Molecular systems biology16, 3 (2020), MSB199083

    2020

  25. [25]

    Michael Schubert, Bertram Klinger, Martina Klünemann, Anja Sieber, Florian Uhlitz, Sascha Sauer, Mathew J Garnett, Nils Blüthgen, and Julio Saez-Rodriguez

  26. [26]

    Perturbation-response genes reveal signaling footprints in cancer gene expression.Nature communications9, 1 (2018), 20

    2018

  27. [27]

    Ali Shojaie, Alexandra Jauhiainen, Michael Kallitsis, and George Michailidis. 2014. Inferring regulatory networks by combining perturbation screens and steady state gene expression profiles.PloS one9, 2 (2014), e82393

    2014

  28. [28]

    Damian Szklarczyk, Annika L Gable, David Lyon, Alexander Junge, Stefan Wyder, Jaime Huerta-Cepas, Milan Simonovic, Nadezhda T Doncheva, John H Morris, Peer Bork, et al. 2019. STRING v11: protein–protein association networks with in- creased coverage, supporting functional discovery in genome-wide experimental datasets.Nucleic acids research47, D1 (2019), ...

    2019

  29. [29]

    Zijia Tang, Minghao Zhou, Kai Zhang, and Qianqian Song. 2024. Scperb: Predict single-cell perturbation via style transfer-based variational autoencoder. Journal of Advanced Research

    2024

  30. [30]

    Allison N Tegge, Charles W Caldwell, and Dong Xu. 2012. Pathway correlation profile of gene-gene co-expression for identifying pathway perturbation.PloS one7, 12 (2012), e52127

    2012

  31. [31]

    The Gene Ontology Consortium. 2021. The Gene Ontology resource: enriching a GOld mine.Nucleic acids research49, D1 (2021), D325–D334

    2021

  32. [32]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    Xiajie Wei, Jiayi Dong, and Fei Wang. 2022. scPreGAN, a deep generative model for predicting the response of single-cell expression to perturbation.Bioinfor- matics38, 13 (2022), 3377–3384

    2022

  34. [34]

    Quin F Wills, Kenneth J Livak, Alex J Tipping, Tariq Enver, Andrew J Goldson, Darren W Sexton, and Chris Holmes. 2013. Single-cell gene expression anal- ysis reveals genetic associations masked in whole-tissue experiments.Nature biotechnology31, 8 (2013), 748–752

    2013

  35. [35]

    Xinyu Yuan, Xixian Liu, Ya Shi Zhang, Zuobai Zhang, Hongyu Guo, and Jian Tang

  36. [36]

    arXiv preprint arXiv:2602.19685

    PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling. arXiv preprint arXiv:2602.19685

    work page arXiv

  37. [37]

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

    2019