Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Diversity by Design: Addressing Mode Collapse Improves scRNA-seq Perturbation Modeling on Well-Calibrated Metrics

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims the trivial mean baseline that often beats deep-learning perturbation models is an artifact of control-referenced, unweighted metrics, and shows that DEG-aware metrics referenced to all perturbed cells sink the mean to…

desk verdict A solid metrics paper: the mean-baseline anomaly is real and the proposed DEG-aware metrics are useful, but the claim that the shared shift is pure artifact is not proven—it measures perturbation-specific signal, not necessarily the whole response. read the letter →

arxiv 2506.22641 v1 pith:7PRFUTOY submitted 2025-06-27 q-bio.GN cs.LGq-bio.MNstat.ML

Gabriel M. Mejia , Henry E. Miller , Francis J. A. Leblanc , Bo Wang , Brendan Swain , Lucas Paulo de Lima Camillo This is my paper · ORCID
classification q-bio.GNcs.LGq-bio.MNstat.ML
keywords single-cellperturbationpredictionmodecollapsecontrolbiasDEG-awaremetricsweightedmeansquarederrordeltaR2scRNA-seqbenchmarkingtechnicalduplicatebaseline
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

Recent single-cell perturbation benchmarks have a strange result: predicting the average of all perturbed cells, with no regard for which perturbation was applied, often beats sophisticated deep-learning models. This paper argues that the result comes from how performance is measured, not from biology. Control-referenced deltas make the shared shift between control and perturbed cells look like perturbation-specific signal, and unweighted MSE is swamped by thousands of unchanged genes, so a collapsed mean prediction scores well. The paper introduces DEG-aware metrics, WMSE and weighted $\Delta$-$R^2$, that use the mean of all perturbed cells as the reference and weight genes by how specifically each perturbation changes them; it also adds negative baselines and a technical-duplicate ceiling to give the metrics a scale. With these measures the mean baseline drops to null performance, and training a model with WMSE instead of plain MSE reduces mode collapse and improves predictions, especially for perturbations with strong signals.

What carries the argument

DEG-aware weights with an all-perturbation reference. For each perturbation, weights are built from absolute t-scores of genes tested against all other perturbed cells (not against control), then min-max normalized, squared, and normalized to sum to one. The metrics are $\mathrm{WMSE} = \sum_i w_i(\mu^p_i - \hat{\mu}^p_i)^2$ and $R^2_w(\Delta) = 1 - \sum_i w_i(\Delta_i - \hat{\Delta}_i)^2 / \sum_i w_i(\Delta_i - \bar{\Delta}_w)^2$, with $\Delta_i = \mu^p_i - \mu^{\mathrm{all}}_i$ and $\hat{\Delta}_i = \hat{\mu}^p_i - \mu^{\mathrm{all}}_i$. These choices concentrate evaluation on the genes that make a perturbation unique, and the constant prediction $\hat{\mu}^p = \mu^{\mathrm{all}}$ is provably scored at or below zero, so mode collapse receives no credit. A technical-duplicate baseline, made by splitting each perturbation's cells in half and predicting one half from the other, calibrates the upper end of the scale by measuring the experiment's own sampling noise floor.

What would settle it

Run the same perturbation panel under a condition where controls are randomized across wells and batches so that $\beta \approx 0$; if the mean baseline still matches fitted models on Pearson($\Delta$), control bias is not the whole story. Alternatively, measure the common shift $\mu_{\mathrm{all}} - \mu_c$ in independent replicate experiments of the same perturbations: if the shift is reproducible across replicates and protocols, it is biological signal, and the all-perturbation reference would remove a real response component.

Watch

Extended reading notes

Core claim

The central claim is that the apparent success of mean predictors in scRNA-seq perturbation benchmarks is a metric artifact rather than a sign that perturbation responses are trivial. The mechanism is a reference-shift correlation: when the control population sits systematically apart from all perturbed populations, the global difference $\Delta_{\mathrm{all}} = \mu_{\mathrm{all}} - \mu_c$ tracks each perturbation-specific difference $\Delta_p = \mu_p - \mu_c$, so Pearson($\Delta$) gives high marks to any predictor that reproduces the shared control-versus-perturbed shift. Unweighted MSE fails in the complementary way, because most genes are unchanged in any given perturbation, so errors on uninformative genes dilute the error signal and hide mode collapse. The paper's positive proposal is to reference all deltas to the mean of all perturbed cells, weight genes by their t-scores against the other perturbations, and calibrate the resulting WMSE and $R^2_w(\Delta)$ with negative baselines and a technical-duplicate ceiling. Under this scheme the mean baseline has strictly negative $R^2_w(\Delta)$ for any perturbation, and retraining a model with WMSE as its loss restores perturbation-specific variance rather than collapsing toward the dataset mean.

Load-bearing premise

The shared expression shift between control and all perturbed cells is experimental bias to be removed, not a genuine biological component that a good perturbation model should be rewarded for predicting.

Editorial extensions

If this is right

  • Scores reported in earlier benchmarks that placed the mean baseline near the top should be reinterpreted as measuring reference bias rather than model skill; rankings on control-referenced Pearson($\Delta$) are unreliable unless the control is shown to be unbiased.
  • Under the proposed metrics a model must predict perturbation-specific DEG signal to score above zero, since the mean baseline's $R^2_w(\Delta)$ is strictly negative for every perturbation.
  • WMSE is not only an evaluation metric but a drop-in training loss: on the Norman19 combination task it cut test-set WMSE by a factor of 2.5 and lifted median $R^2_w(\Delta)$ from $-0.068$ to $0.555$ for an established perturbation model, and on the Replogle22 unseen-gene task it improved every weighted metric, most clearly for perturbations with strong DEG signals.
  • A technical-duplicate baseline gives the field a principled performance ceiling: a model performing at that level predicts mean effects as accurately as the experiment's own sampling noise allows.
  • Switching the delta reference from control cells to all perturbed cells also changes DEG calling, removing the shared control-versus-perturbed genes so that DEG lists mark what is unique to each perturbation rather than what is common to all.

Reading between the lines

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

  • If the shared control-to-perturbed shift is partly biological, such as a conserved stress or growth program, then treating it as bias would discard a real response component; this can be tested by checking whether $\mu_{\mathrm{all}} - \mu_c$ reproduces across independent replicate experiments and different control preparations.
  • The reference-shift mechanism should also inflate other control-referenced scores, including top-k DEG overlap and cosine distances from control, so the all-perturbation reference is likely to generalize beyond WMSE and $R^2_w(\Delta)$.
  • A model architecture that explicitly splits each prediction into a shared response component plus a perturbation-specific residual, trained with WMSE, is the natural design consequence; the paper's variance-recovery results suggest such a decomposition would fit the corrected objective.
  • An outer-loop version of the method that computes DEG weights on held-out splits or separate replicate experiments would confirm that the reported training gains come from the weighting itself rather than from information shared with the evaluation set.
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 / 6 minor

Summary. This paper addresses the observation that predicting the mean of all perturbed cells (μ_all) often outperforms sophisticated scRNA-seq perturbation-response models. The authors argue that this is a metric artifact caused by control-referenced deltas and unweighted error aggregation: when control cells are systematically shifted from all perturbed cells, the shared reference shift (denoted β) dominates perturbation-specific differences. They support this with negative-binomial simulations and two real Perturb-seq datasets, then propose DEG-aware weighted metrics (WMSE and R2w(Δ)) referenced to μ_all, a technical-duplicate positive baseline, and a demonstration that retraining GEARS with WMSE loss improves performance on the proposed metrics. Code is provided.

Significance. If the premise is correct, the paper makes a useful methodological contribution: it offers a plausible explanation for a puzzling benchmark result, introduces metrics with a clear null baseline and an empirical ceiling, and proposes a simple training modification to reduce mode collapse. The technical-duplicate baseline is a practical calibration device, the proof that constant predictions score at most zero under R2w(Δ) is a clean property, and the code release supports reproducibility. However, the central premise—that the shared control-to-perturbation shift is a technical artifact rather than a genuine shared biological response—is assumed rather than demonstrated, and the evaluation of WMSE-trained models is partly circular with the proposed weighting. If the premise is strengthened or the contribution is reframed as measuring perturbation-specific deviations from the average perturbation, the paper would be a solid contribution to benchmarking methodology.

major comments (3)
  1. [§2.3, §3.3.1, §4.2] The load-bearing claim that the shared shift between control and all perturbed cells is a technical artifact is not established. The observation that many DEGs are shared across perturbations (Fig. 2A, Suppl. Fig. 5A) is equally consistent with a genuine common response program such as a stress or growth-arrest module. The simulations define β as an injected artifact (Eq. 2), so they cannot validate this premise on real data, and the real-data β experiment in Fig. 2D–E only interpolates μ_c toward μ_all, making the reported Pearson(Δ) trend a definitional consequence rather than independent evidence. Because R2w(Δ) uses μ_all as reference and Appendix B shows a constant prediction scores at most zero, a model that correctly predicts a real shared response component is penalized. Please either provide evidence that the shared component is technical (e.g., association with batch or QC covariates, or annotation of shared DEGs against known common response programs) or reframe the contribution as metrics for perturbation-specific deviations from the average perturbation.
  2. [§4.5, Fig. 4, Suppl. Fig. 8] The claim that WMSE training reduces mode collapse is evaluated primarily on WMSE and DEG-weighted R2w(Δ), which use the same DEG-vs-Rest statistics that define the WMSE loss. The Pearson(Δ) result in Fig. 4C still filters to the same DEG set, so it does not fully break the circularity. Please report the training comparison on a metric that does not depend on the proposed weighting, such as unweighted all-gene R2(Δ) with the μ_all reference, or justify that the observed gains are not an artifact of optimizing the evaluation objective. Complementing Fig. 4D with a quantitative variance-matching statistic would also help.
  3. [§3.3.1, Eq. (6)] Using μ_all as the delta reference makes the proposed metrics depend on the composition of the perturbation panel. Adding or removing perturbations changes μ_all, and hence changes Δ, R2w(Δ), and the DEG-vs-Rest weights for every perturbation. The paper discusses the effect of k on the older Pearson(Δ) metric (Appendix D) but does not analyze the stability of the proposed metrics under perturbation-panel subsampling. Since the metrics are intended for benchmarking across datasets with very different panel sizes (175 versus 1334 perturbations here), please add a stability analysis (e.g., subsampling perturbations and recomputing R2w(Δ) and WMSE) or explicitly discuss this dependence.
minor comments (6)
  1. [§3.3.1 and Appendix B] The text says the mean baseline yields a 'strictly negative' R2w(Δ), but the proof in Appendix B establishes only R2w(Δ) ≤ 0, with equality possible when the weighted mean of Δ is zero. Please correct the statement in both places.
  2. [§3.3.3] 'sc.tl.ran_genes_groups' should be 'sc.tl.rank_genes_groups'.
  3. [§4.1 and Table 2] The statement that increasing the number of genes g reduced 'all error metrics' is too broad; in Table 2, Pearson(Δ) correlations with g are near zero for all gene sets. Please restrict the claim to MSE-type metrics.
  4. [§4.5 and §3.3.3] Please clarify whether the weights used to evaluate held-out perturbations are computed from the held-out cells themselves or from training cells, as this affects how the WMSE-training gains should be interpreted.
  5. [Abstract] 'shared reference shifts, not genuine biological change, drives high performance' should be 'drive high performance' for subject-verb agreement.
  6. [§3.3.3, step (iv)] The square in the weight definition is an arbitrary tuning choice; please provide a sensitivity analysis or a justification for this exponent, since it changes how strongly a small set of genes dominates the metric.

Circularity Check

3 steps flagged · score 4.0 of 10

The control-bias conclusion is partly baked into the simulation and metric definitions; WMSE gains are partially tautological with the evaluation metrics.

  1. fitted input called prediction [Sec. 3.1 (simulation definition of β) and Appendix A (real-data β experiment)]
    "Following reasonable priors, we define λi = µall i − µc i as the difference between the average perturbed expression and the control expression in the Norman19 dataset. ... β a global dataset parameter controlling the severity to which λi is applied (zero for a perfectly centered control). ... We then created synthetic control data by interpolating in equivalent steps of 0.1∆ between µall (0∆) and µc (1∆)."

    The quantity whose effect is being measured, βλ, is defined as the shared offset µall−µc. In the generative model, µp = l α (µc + βλ) while µall ≈ µc + βλ, so ∆all ≈ βλ and ∆p ≈ βλ + (α−1)(µc+βλ). Pearson(∆p,∆all) therefore increases with β by construction rather than by empirical discovery. The real-data β experiment performs the same interpolation, moving the control mean toward µall and algebraically guaranteeing a higher mean-baseline correlation. Thus the paper's conclusion that 'shared reference shifts, not genuine biological change' drive mean-baseline performance is an assumption embedded in how control bias is defined, not a prediction independently confirmed by these experiments.

  2. self definitional [Sec. 3.3.1 and Appendix B (R2_w(∆) definition and derivation)]
    "Because of the properties of R2, any constant average predictions (µ̂p = µall) will yield a strictly negative result for any specific perturbation (see Appendix B for derivation)."

    Equation (6) defines R2_w with ∆i = µp_i − µall_i and ∆̂i = µ̂p_i − µall_i; setting µ̂p = µall makes ∆̂ = 0, and Appendix B proves R2_w = 1 − Σ wi ∆i² / (Σ wi ∆i² − ∆̄w²) ≤ 0. The 'null performance' of the mean baseline is therefore a designed property of the metric, not an empirical result. The paper discloses this explicitly, so the circularity is mild, but the abstract's phrasing 'the mean baseline sinks to null performance' is a definitional consequence rather than independent evidence that the earlier mean-baseline success was artifactual.

1 more flagged steps
  1. fitted input called prediction [Sec. 4.5 (MSE vs WMSE training evaluation)]
    "Switching the training objective from MSE to WMSE reduces the test-set WMSE by a factor of 2.5 and lifts the median R2w(∆) from −0.068 to 0.555, bringing GEARS to within striking distance of the technical duplicate baseline. The same trend appears in Pearson(∆) computed after filtering to DEGs (vs Rest) (Fig. 4C), a metric whose definition contains no weighting and thus rules out information leakage from the training weights."

    GEARS is retrained with WMSE, which is the same weighted, DEG-vs-Rest objective used to define the WMSE and R2_w evaluation metrics (Secs. 3.3.2–3.3.3). A model optimized for WMSE will tend to improve on WMSE and on R2_w (whose numerator is the same weighted squared error) for reasons that are partly tautological, though held-out test sets mean the improvement still requires generalization. The metric offered as independent, Pearson on DEGs vs Rest, nevertheless uses the same DEG-vs-Rest reference and gene prioritization as the loss weights, so it is not fully insulated from the training objective. The claim that WMSE training improves model performance is therefore partially a fitted-input-called-prediction.

full rationale

This is a metrics-development paper, and it is transparent that R2_w gives the mean baseline a null/negative score by construction (Appendix B). The more substantive circularity concern is the causal claim that shared reference shifts are artifacts: the simulation defines the control-bias parameter β through λ = µall−µc, and the real-data β experiment interpolates the control mean toward µall. The observed rise in mean-baseline Pearson(∆) is therefore a consequence of the construction, not an independent test against the alternative that the shared shift is a real common biological response. The WMSE-training result is partially circular because the evaluation metrics share the DEG-vs-Rest weighting and µall reference with the training loss, and the 'independent' Pearson-DEGs metric still uses the same DEG prioritization. However, the paper also cites external benchmarks showing mean-baseline success, and the proposed metrics, technical-duplicate baseline, and WMSE loss have independent content as tools. The central interpretation is therefore partly assumed rather than fully derived, but the paper is not empty or wholly self-referential.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new physical or biological entities are introduced; the technical duplicate baseline is a data-derived construct, not an invented entity. The empirical claims rest on domain assumptions about what constitutes artifact versus biology, and on the hand-chosen weighting scheme. The metrics are definitions rather than derivations, so the contribution is a calibrated evaluation framework rather than a first-principles result.

free parameters (1)
  • Weight squaring exponent = 2
    In Section 3.3.3, weights are squared to accentuate differences; the exponent 2 is chosen by hand and directly shapes both the metric and the training loss.
assumptions (5)
  • domain assumption The systematic difference between control and perturbed cells is an artifact ('control bias') rather than a biological signal to be rewarded.
    Introduced in Section 2.3 and used throughout, this justifies replacing the control reference with the mean of all perturbed cells in Sections 3.3.1 and 3.3.3.
  • domain assumption The mean of all perturbed cells (mu_all) is a bias-free reference for delta computation and DEG identification.
    Section 3.3.1 defines all deltas against mu_all and Section 3.3.3 computes DEGs vs Rest; this assumes mu_all is a meaningful, stable center that does not introduce its own bias.
  • domain assumption t-test scores computed against all other perturbations provide reliable gene importance weights even with as few as 64 cells per perturbation.
    Sections 3.2 and 3.3.3 rely on scanpy rank_genes_groups with t-test_overestim_var and reference='rest'; low cell counts make variance estimates noisy, and the choice of test could change which genes are prioritized.
  • domain assumption The simulation priors (lambda_i, theta_i, mu_c, sigma_l) estimated from Norman19 are representative of real perturbation datasets.
    Section 3.1 sets lambda_i = mu_all - mu_c from Norman19 and estimates other parameters from the same data, anchoring the simulation to the characteristics of a single dataset.
  • standard math Standard properties of weighted sums and R2 are used in the proof that constant predictions yield non-positive R2_w(delta).
    Appendix B uses the identity sum w_i (Delta_i - Delta_bar)^2 = sum w_i Delta_i^2 - Delta_bar^2 for weights summing to 1; no new mathematics is introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diversity by Design: Addressing Mode Collapse Improves scRNA-seq Perturbation Modeling on Well-Calibrated Metrics." pith.science (2026). https://pith.science/paper/7PRFUTOY

@misc{pith2026250622641,
  author       = {Pith},
  title        = {Pith review of: Diversity by Design: Addressing Mode Collapse Improves scRNA-seq Perturbation Modeling on Well-Calibrated Metrics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PRFUTOY}},
  note         = {Machine review of arXiv:2506.22641}
}
abstract

Recent benchmarks reveal that models for single-cell perturbation response are often outperformed by simply predicting the dataset mean. We trace this anomaly to a metric artifact: control-referenced deltas and unweighted error metrics reward mode collapse whenever the control is biased or the biological signal is sparse. Large-scale \textit{in silico} simulations and analysis of two real-world perturbation datasets confirm that shared reference shifts, not genuine biological change, drives high performance in these evaluations. We introduce differentially expressed gene (DEG)-aware metrics, weighted mean-squared error (WMSE) and weighted delta $R^{2}$ ($R^{2}_{w}(\Delta)$) with respect to all perturbations, that measure error in niche signals with high sensitivity. We further introduce negative and positive performance baselines to calibrate these metrics. With these improvements, the mean baseline sinks to null performance while genuine predictors are correctly rewarded. Finally, we show that using WMSE as a loss function reduces mode collapse and improves model performance.

Figures

Figures reproduced from arXiv: 2506.22641 by the authors.

Figure 1
Figure 1. (A) Under increasing amounts of systematic control bias, the Pearson correlation between ∆p and ∆all increases artificially. (B) True biological signal is diluted in scRNA-seq data causing mode collapse in model predictions. Introducing biologically aware weights (WMSE) ameliorates this problem. (C) We intro￾duce metrics with a performance scale calibrated by negative and positive baselines. We also find that WMSE a… view at source ↗
Figure 2
Figure 2. (A) Trinary (up, down, or unchanged) clustermap of significant DEGs for each perturbation compared to control. (B) Schematic showing high correlation between ∆p and ∆all due to the systematic bias of µ c . (C) Pearson(∆) performance of the mean baseline (µ all) under increasing control bias (β) in simulations. Trend line shows moving average. (D) Pearson(∆) performance of µ all under increasing β in the Norman19 dat… view at source ↗
Figure 3
Figure 3. Weighted MSE is sensitive to differences in niche per￾turbation signals. (A) Violin plot showing MSE between true perturbation profiles µ p and two possible predictions: the mean baseline µ all and a modified version µ all∗ which sets the top 25 DEGs to be perfectly predicted. (B) Same as (A) but MSE was weighted by the normalized DEG score (WMSE). (C-D) Same as (A-B) but for the Replogle22 dataset. 4.4. Improved Me… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: DEG score-weighted loss reduces mode collapse and improves model performance (Norman19). (A) WMSE between prediction and ground-truth perturbation mean. X labels: µ c (con￾trol mean), µ all (mean of all perturbed cells), predictions from GEARS model with MSE or WMSE lo…
Figure 5
Figure 5. Figure 5: Supplemental figure accompanying [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: DEG-aware metrics, elimination of control bias, and addition of negative and positive baselines provide greater sensitivity and calibration to assess perturbation response model performance. (A) In the Norman19 dataset, MSE between ground truth and baselines or model p…
Figure 7
Figure 7. Figure 7: Effect of perturbation strength (measured by number of significant DEGs) on R 2 w(∆p , ∆ˆ p ) metric in technical duplicate baseline. (A) Scatter plots showing correlation between perturbation effects (∆ = µ p − µ all) when using the first half of the data to predict t…
Figure 8
Figure 8. Figure 8: Supplemental figure accompanying [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: DEG score-weighted loss improves model performance on unseen gene prediction task, especially for stronger perturbations. (A) Performance of baselines and GEARS perturbation prediction (with MSE or WMSE loss), compared using WMSE metric vs ground truth perturbation mea…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. 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

26 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

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

    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, 2024

    work page 2024

  3. [3]

    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, 2024

    arXiv

  4. [4]

    Z., and Szalai, B

    Csendes, G., Sanz, G., Szalay, K. Z., and Szalai, B. Benchmarking foundation cell models for post-perturbation rna-seq prediction. BMC genomics, 26 0 (1): 0 393, 2025

    work page 2025

  5. [5]

    scgpt: toward building a foundation model for single-cell multi-omics using generative ai

    Cui, H., Wang, C., Maan, H., Pang, K., Luo, F., Duan, N., and Wang, B. scgpt: toward building a foundation model for single-cell multi-omics using generative ai. Nature Methods, 21 0 (8): 0 1470--1480, 2024

    work page 2024

  6. [6]

    P., Jerby-Arnon, L., Marjanovic, N

    Dixit, A., Parnas, O., Li, B., Chen, J., Fulco, C. P., Jerby-Arnon, L., Marjanovic, N. D., Dionne, D., Burks, T., Raychowdhury, R., et al. Perturb-seq: dissecting molecular circuits with scalable single-cell rna profiling of pooled genetic screens. cell, 167 0 (7): 0 1853--1866, 2016

    work page 2016

  7. [7]

    xtrimogene: an efficient and scalable representation learner for single-cell rna-seq data

    Gong, J., Hao, M., Cheng, X., Zeng, X., Liu, C., Ma, J., Zhang, X., Wang, T., and Song, L. xtrimogene: an efficient and scalable representation learner for single-cell rna-seq data. Advances in Neural Information Processing Systems, 36: 0 69391--69403, 2023

    work page 2023

  8. [8]

    Large-scale foundation model on single-cell transcriptomics

    Hao, M., Gong, J., Zeng, X., Liu, C., Guo, Y., Cheng, X., Wang, T., Ma, J., Zhang, X., and Song, L. Large-scale foundation model on single-cell transcriptomics. Nature methods, 21 0 (8): 0 1481--1491, 2024

    work page 2024

Show all 26 references
  1. [9]

    scgenept: Is language all you need for modeling single-cell perturbations? bioRxiv, pp.\ 2024--10, 2024

    Istrate, A.-M., Li, D., and Karaletsos, T. scgenept: Is language all you need for modeling single-cell perturbations? bioRxiv, pp.\ 2024--10, 2024

    2024

  2. [10]

    M., Jindal, K., Solnica-Krezel, L., and Morris, S

    Kamimoto, K., Stringa, B., Hoffmann, C. M., Jindal, K., Solnica-Krezel, L., and Morris, S. A. Dissecting cell identity via network inference and in silico gene perturbation. Nature, 614 0 (7949): 0 742--751, 2023

    2023

  3. [11]

    Benchmarking ai models for in silico gene perturbation of cells

    Li, C., Gao, H., She, Y., Bian, H., Chen, Q., Liu, K., Wei, L., and Zhang, X. Benchmarking ai models for in silico gene perturbation of cells. bioRxiv, pp.\ 2024--12, 2024 a

    2024

  4. [12]

    A systematic comparison of single-cell perturbation response prediction models

    Li, L., You, Y., Liao, W., Fan, X., Lu, S., Cao, Y., Li, B., Ren, W., Fu, Y., Kong, J., et al. A systematic comparison of single-cell perturbation response prediction models. bioRxiv, pp.\ 2024--12, 2024 b

    2024

  5. [13]

    B., Jordan, M

    Lopez, R., Regier, J., Cole, M. B., Jordan, M. I., and Yosef, N. Deep generative modeling for single-cell transcriptomics. Nature methods, 15 0 (12): 0 1053--1058, 2018

    2018

  6. [14]

    A., and Theis, F

    Lotfollahi, M., Wolf, F. A., and Theis, F. J. scgen predicts single-cell perturbation responses. Nature methods, 16 0 (8): 0 715--721, 2019

    2019

  7. [15]

    L., Srivatsan, S

    Lotfollahi, M., Klimovskaia Susmelj, A., De Donno, C., Hetzel, L., Ji, Y., Ibarra, I. L., Srivatsan, S. R., Naghipourfar, M., Daza, R. M., Martin, B., et al. Predicting cellular responses to complex perturbations in high-throughput screens. Molecular systems biology, 19 0 (6):...

    2023

  8. [16]

    M., Pogson, A

    Nadig, A., Replogle, J. M., Pogson, A. N., Murthy, M., McCarroll, S. A., Weissman, J. S., Robinson, E. B., and O’Connor, L. J. Transcriptome-wide analysis of differential expression in perturbation atlases. Nature Genetics, pp.\ 1--10, 2025

    2025

  9. [17]

    M., Horlbeck, M

    Norman, T. M., Horlbeck, M. A., Replogle, J. M., Ge, A. Y., Xu, A., Jost, M., Gilbert, L. A., and Weissman, J. S. Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science, 365 0 (6455): 0 786--793, 2019

    2019

  10. [18]

    D., Shen, C., Gross, T., Min, J., Garda, S., Yuan, B., Schumacher, L

    Peidli, S., Green, T. D., Shen, C., Gross, T., Min, J., Garda, S., Yuan, B., Schumacher, L. J., Taylor-King, J. P., Marks, D. S., et al. scperturb: harmonized single-cell perturbation data. Nature Methods, 21 0 (3): 0 531--540, 2024

    2024

  11. [19]

    M., Saunders, R

    Replogle, J. M., Saunders, R. A., Pogson, A. N., Hussmann, J. A., Lenail, A., Guna, A., Mascibroda, L., Wagner, E. J., Adelman, K., Lithwick-Yanai, G., et al. Mapping information-rich genotype-phenotype landscapes with genome-scale perturb-seq. Cell, 185 0 (14): 0 2559--2575, 2022

    2022

  12. [20]

    Predicting transcriptional outcomes of novel multigene perturbations with gears

    Roohani, Y., Huang, K., and Leskovec, J. Predicting transcriptional outcomes of novel multigene perturbations with gears. Nature Biotechnology, 42 0 (6): 0 927--935, 2024

    2024

  13. [21]

    S., Quake, S

    Rosen, Y., Roohani, Y., Agarwal, A., Samotor c an, L., Consortium, T. S., Quake, S. R., and Leskovec, J. Universal cell embeddings: A foundation model for cell biology. bioRxiv, pp.\ 2023--11, 2023

    2023

  14. [22]

    scperb: Predict single-cell perturbation via style transfer-based variational autoencoder

    Tang, Z., Zhou, M., Zhang, K., and Song, Q. scperb: Predict single-cell perturbation via style transfer-based variational autoencoder. Journal of Advanced Research, 2024

    2024

  15. [23]

    V., Xiao, L., Chopra, A., Chaffin, M

    Theodoris, C. V., Xiao, L., Chopra, A., Chaffin, M. D., Al Sayed, Z. R., Hill, M. C., Mantineo, H., Brydon, E. M., Zeng, Z., Liu, X. S., et al. Transfer learning enables predictions in network biology. Nature, 618 0 (7965): 0 616--624, 2023

    2023

  16. [24]

    Perteval-scfm: Benchmarking single-cell foundation models for perturbation effect prediction

    Wenteler, A., Occhetta, M., Branson, N., Huebner, M., Curean, V., Dee, W., Connell, W., Hawkins-Hooker, A., Chung, P., Ektefaie, Y., et al. Perteval-scfm: Benchmarking single-cell foundation models for perturbation effect prediction. bioRxiv, pp.\ 2024--10, 2024

    2024

  17. [25]

    A., Angerer, P., and Theis, F

    Wolf, F. A., Angerer, P., and Theis, F. J. Scanpy: large-scale single-cell gene expression data analysis. Genome biology, 19: 0 1--5, 2018

    2018

  18. [26]

    M., Nassar, M., Osi \'n ski, B., Eksi, R., Zhang, K., and Graepel, T

    Wu, Y., Wershof, E., Schmon, S. M., Nassar, M., Osi \'n ski, B., Eksi, R., Zhang, K., and Graepel, T. Perturbench: Benchmarking machine learning models for cellular perturbation analysis. arXiv preprint arXiv:2408.10609, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.