Pith. sign in

REVIEW 4 major objections 6 minor 23 references

The paper claims that transcriptional responses to unseen genetic perturbations can be predicted accurately by conditioning on a sparse, perturbation-specific subgraph of a knowledge graph and by explicitly separating signal from noise duri

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 21:50 UTC pith:QFCLLEX2

load-bearing objection A genuinely new mechanism for sparsity in perturbation prediction, with real K562 gains, but the 'consistently best' claim overreaches and the DEG-aware metrics are partly aligned with the training signal. the 4 major comments →

arxiv 2602.18885 v2 pith:QFCLLEX2 submitted 2026-02-21 cs.CE

Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction

Yinhua Piao , Hyomin Kim , Seonghwan Kim , Yunhak Oh , Junhyeok Jeon , Sang-Yeon Hwang , Jaechang Lim , Woo Youn Kim
show 2 more authors
Chanyoung Park Sungsoo Ahn
This is my paper
classification cs.CE
keywords transcriptional response predictiongenetic perturbationsmean collapsedifferentially expressed genessingle-cell RNA-seqknowledge graphsubgraph selectionadaptive regularization
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.

AdaPert is a framework for predicting how gene expression changes after a genetic perturbation that was not seen in training. The paper's thesis is that a common failure — mean-collapse, where models predict the average response and miss a perturbation's true target genes — is not a data problem but a design problem, and can be fixed with two moves: extract a sparse subgraph of a biological knowledge graph that is specific to the perturbed gene, and train with adaptive losses that suppress changes in non-responsive genes while aligning the subgraph representation to the differentially expressed genes. On two large single-cell CRISPR datasets, AdaPert outperforms prior approaches, with the largest gains on DEG-aware metrics such as recovery and ranking of differentially expressed genes. If this holds, perturbation models become substantially more trustworthy for ranking genes and directions of change, which matters for functional genomics and drug-target discovery.

Core claim

The central discovery is that mean-collapse in perturbation response prediction can be reduced by making both graph context and supervision perturbation-specific. Standard models pass the perturbed gene through a dense, static knowledge graph embedding, and train with a mean-squared reconstruction loss that is dominated by the many non-responsive genes; this drives predictions toward the global average. AdaPert instead scores every graph node against a text-derived embedding of the perturbed gene, applies differentiable Gumbel-Softmax sampling to select a small subgraph, and uses that subgraph's summed embeddings as the perturbation context. Training then combines a global reconstruction los

What carries the argument

Perturbation-conditioned subgraph extraction: a language model embeds the perturbed gene's textual description, each node in a protein-interaction knowledge graph is scored by an MLP combining its structural embedding with that semantic embedding, and Gumbel-Softmax sampling selects a sparse node set, summarized by summing selected node embeddings. This subgraph is trained with an adaptive objective consisting of three losses: global reconstruction, a Huber penalty on predicted changes of non-DEG genes, and a cosine alignment between the subgraph representation and a response target made from DEG effects.

Load-bearing premise

The claim presupposes that the DEG/non-DEG partition computed from training-data p-values sensibly separates true signal from noise; if those labels are noisy or statistically underpowered, the adaptive losses will suppress real effects and the reported DEG-aware gains may be artifacts of the supervision.

What would settle it

Re-run the same training pipeline with DEG masks computed from an independent, higher-powered call on held-out perturbation data, then compare DES@50 with the paper's masks; if the DEG-aware advantage disappears or reverses, the gain was driven by the choice of training labels rather than by the subgraph context.

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

If this is right

  • Unseen-perturbation predictions will rank differentially expressed genes more accurately, so high-confidence predictions can be prioritized for experimental follow-up.
  • The non-DEG suppression and context alignment act as explicit anti-mean-collapse mechanisms, so the gains should concentrate in small-effect perturbations, which are exactly those where the average response is most misleading.
  • Because the subgraph is conditioned on the perturbed gene, the model can give a per-perturbation list of network-relevant genes, improving interpretability over dense global embeddings.
  • Pathway-level enrichment of predicted responses correlates with ground truth, suggesting the model preserves coordinated regulatory signatures, not just individual gene statistics.

Where Pith is reading between the lines

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

  • Extending beyond the paper: the same adaptive-loss recipe could be applied to any autoencoder-based perturbation predictor, with or without a knowledge graph, by masking non-responsive genes during training.
  • The method's reliance on training-derived DEG labels suggests an extension: iteratively refine or self-train the masks to avoid entrenching false negatives from underpowered single-cell data.
  • The sparse-subgraph readout is a simple sum; a natural next step would be attention-weighted aggregation or multiple subgraphs to capture several pathways of response.
  • The framework should transfer to other condition-specific predictions, such as drug responses or combinatorial perturbations, whenever a per-condition relational subgraph can be defined.

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

4 major / 6 minor

Summary. The paper proposes AdaPert, a perturbation-conditioned framework for predicting transcriptional responses to unseen genetic perturbations. AdaPert extracts a sparse, perturbation-specific subgraph from a biological knowledge graph via differentiable Gumbel-Softmax node selection and applies an adaptive learning scheme with three losses: a global reconstruction loss, a Huber loss that suppresses predicted changes on genes labeled as non-DEGs, and an alignment loss that matches the selected subgraph representation to a DEG-masked response vector. The authors evaluate on the K562.Replogle and RPE1.Replogle datasets under an unseen-perturbation split, reporting gains over baselines such as TxPert and MorPH on global metrics and DEG-aware metrics, with particular emphasis on reduced mean-collapse and improved DEG recovery.

Significance. If the reported results are robust, AdaPert makes a useful contribution to perturbation-response prediction by replacing dense, static knowledge-graph priors with adaptive sparse subgraph selection and by explicitly separating signal from noise in a way that mitigates mean-collapse. The work has concrete strengths: it evaluates perturbation discrimination with PDS in addition to correlation, provides an effect-size-stratified analysis, includes an ablation separating the non-DEG loss and the context module, and offers a pathway-level validation. These design choices go beyond simple benchmark reporting. However, the headline claim of consistent, substantial improvement is weakened by the RPE1 numbers, particularly the PDS score, and the DEG-aware metrics are partially aligned with the training supervision, so additional evidence is needed to establish that the reported DEG gains reflect genuine biological signal recovery rather than threshold-matching artifacts.

major comments (4)
  1. [§5.1, Table 1] The text states that AdaPert achieves the best performance on both datasets, but Table 1 shows MorPH has higher PDS on RPE1 (0.688 vs. 0.663). The K562 PDS gain is large and meaningful, but on RPE1 the Pearson-Δ margin over TxPert (0.674 vs. 0.655) is also small. The claim of consistent superiority should be qualified per metric/dataset, or supported with significance tests or confidence intervals.
  2. [§3.4 (Eqs. 14–21) and §B.2 (Eq. 27)] The DEG-aware evaluation is partially aligned with the training signal. The sets D(p) and complement used for L_non and L_align are defined from training p<0.05 calls, while DES and the DE metrics use the same p-value/FDR threshold on test perturbations. Because test perturbations are unseen, this is not flatly circular, but the reported DES improvements could partly reflect the model being optimized to reproduce the statistical threshold's noise structure. To make the central claim load-bearing, please evaluate with alternative DEG definitions (e.g., different FDR thresholds, absolute effect-size cutoffs, or an independent DEG-calling method) and report whether the DEG-aware gains persist.
  3. [Appendix F, Table 10] The RPE1 DEG-aware advantages over TxPert are numerically marginal: DES@50 is 0.244 vs. 0.242 and DES@100 is 0.320 vs. 0.314. The text nonetheless claims AdaPert 'achieves the best performance across all reported metrics' and shows 'consistent improvements.' Without error bars or significance testing, these differences are within likely noise, and the claim of substantial DEG-aware improvement is not supported on RPE1.
  4. [§3.4, Eq. (17); Appendix E] The Huber threshold δ is stated to be 'set proportional to the empirical standard deviation of non-DEG effects' and fixed across perturbations, but the proportionality constant is not reported. The main free hyperparameters (λ_non, λ_align, τ, T, k, DEG p-value threshold) are not given for the AdaPert model; Appendix Table 8 lists only the TxPert search space. Reporting final hyperparameter values for all datasets is necessary for reproducibility.
minor comments (6)
  1. [Fig. 2] The figure contains garbled labels ('0+0000-00 DEG', 'Differential Expressed Analysis') and inconsistent use of 'DEG' versus 'DE'; these should be cleaned up.
  2. [Table 2 caption] The caption says 'DEG overlap@k' while the metric in §B.2 is 'Differential Expression Score (DES)'. Please unify the terminology.
  3. [Table 3] The table reports means without standard deviations despite the main tables using mean ± std; add the same uncertainty information or state why it is omitted.
  4. [Eq. (10)] The dimensions of the weights w and W_c are not defined; please specify so the node-scoring MLP is unambiguous.
  5. [Appendix A.2, Fig. 9] The description of 'DEG coverage as a function of graph hop distance' is vague. Please define what 'predicted genes' are in this figure and how the coverage numerator/denominator are computed.
  6. [Appendix E] No code repository or model checkpoint is provided. Given the number of hyperparameters and the use of GPT-4o embeddings, a public implementation would strengthen reproducibility.

Circularity Check

1 steps flagged

DEG-aware gains are partly entangled with the training DEG supervision, but the core prediction task is not circular.

specific steps
  1. fitted input called prediction [§3.4 (Eq. 14, 16, 18, 20) and §B.2 (Eq. 27)]
    ""Using the training data, we perform a statistical test for each gene and obtain a p-value q(p)i. We define the DEG and non-DEG sets as D(p)={i|q(p)i<0.05}, D̄(p)={i|q(p)i≥0.05}." ... "we penalize predicted perturbation changes on D̄(p) using a Huber loss" ... "y(p)i = ΔXpi, i∈D(p); 0, i∈D̄(p)" ... "The Differential Expression Score ... DES(p)=|Gtrue(p)∩Gpred(p)|/|Gtrue(p)|""

    The training supervision is built from the same statistical DEG concept used by the headline evaluation. Eq. 14 defines D(p)/D̄(p) from p<0.05; Lnon shrinks predicted changes on D̄(p) toward zero, and Lalign makes the context representation encode only D(p)-masked effects. DES (Eq. 27) then rewards recovery of genes that are significant DEGs on held-out perturbations under the same threshold-style definition. This does not make the test predictions equal to the training labels, since test perturbations are unseen, but it means the reported DES improvements are partially aligned with the training signal rather than being an independent confirmation of biological signal recovery. The metric family is not purely external: the model is explicitly optimized to match the noise structure of a p<0

full rationale

There is no equation that reduces by construction to its inputs, no self-citation chain, and no imported uniqueness theorem. The central derivation is self-contained: AdaPert is trained on control/perturbed expression profiles and evaluated on unseen perturbations. The main concern is the DEG-aware evaluation: the training losses Lnon and Lalign use training-data DEG masks defined by p<0.05 (Eq. 14), while DES and related DEG metrics use the same kind of differential-expression threshold on test data (Eq. 27). This creates a partial statistical alignment between the training objective and the headline metric family, so the DES improvements are not fully independent evidence that true perturbation-specific signal is recovered. However, the test perturbations are unseen, PDS and Pearson-Δ gains do not use DEG labels, the HIRA GSEA is an external functional check, and no fitted parameter is renamed as a prediction. The circularity is therefore partial and metric-specific rather than a collapse of the whole derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method combines known components (GNN message passing, Gumbel-Softmax selection, auxiliary losses) and introduces no new biological entities. The load-bearing free choices are the loss weights and selection threshold, which are underreported. The DEG split is simultaneously the training signal and the evaluation target, which is the main audit point; the HIRA GSEA provides the only outside-the-benchmark validation.

free parameters (7)
  • λ_non (non-DEG loss weight) = not stated; Fig. 5 sweeps 0–0.1
    Weight of the non-DEG Huber loss in Eq. 21; central to the signal/noise separation claim; operating point not reported in text.
  • λ_align (alignment loss weight) = not stated
    Weight of the DEG-alignment loss L_align in Eq. 21; never specified numerically.
  • δ (Huber threshold) = proportional to empirical SD of non-DEG effects; proportion unspecified
    Threshold in Eq. 16–17 fit to data statistics and fixed across perturbations; the proportionality constant is a free choice.
  • τ (Gumbel temperature) = not stated
    Temperature in Eq. 12 controlling hardness of node selection; annealing schedule not described.
  • T (node selection threshold) = not stated
    Threshold on Gumbel scores defining the extracted subgraph G_p (Eq. 12–13); never given.
  • k (STRING top-k edge filtering) = 10 or 20 (Table 6)
    Graph preprocessing choice that determines the template graph and hence the subgraph selection space.
  • DEG p-value threshold = 0.05
    Defines D(p) and non-D(p) (Eq. 14); drives both training losses and the evaluation's notion of signal.
axioms (5)
  • domain assumption STRING PPI edges (top-k filtered) capture functional relationships relevant to perturbation response
    The entire subgraph extraction (§3.3) assumes graph proximity implies transcriptional co-response; Fig. 9 verifies DEG coverage only post hoc on predictions.
  • domain assumption GenePT/GPT-4o semantic embeddings of NCBI gene descriptions are predictive of perturbation-relevant function
    Neighbor scoring (Eq. 10) conditions selection on a projected semantic embedding; validity is inferred only from final benchmark metrics.
  • domain assumption DEG calls at p<0.05 on training data are a correct binary split of signal vs noise
    Eqs. 14–21 build both the suppression loss and the alignment target on this split; noisy or low-power calls propagate directly into the learned predictions.
  • standard math Gumbel-Softmax with hard threshold T is a valid differentiable subgraph selection
    Eq. 12 is a standard reparameterization (Jang et al., 2017), but the hard threshold T is an ad hoc binarization step.
  • domain assumption cell-eval metrics and ground-truth test DEGs are the right measure of biological fidelity
    All conclusions rest on this evaluation contract; different DEG thresholds or metric definitions would change the reported ranking.

pith-pipeline@v1.3.0-alltime-deepseek · 18670 in / 18278 out tokens · 159517 ms · 2026-08-02T21:50:00.501855+00:00 · methodology

0 comments
read the original abstract

Predicting high-dimensional transcriptional responses to genetic perturbations is challenging because signals are sparse and experimental noise is severe. Existing methods often suffer from mean collapse, achieving high correlation by predicting the global average expression rather than perturbation-specific responses, which yields false positives and poor interpretability. Methods that add biological knowledge graphs typically treat them as dense, static priors shared across perturbations, propagating noise. We propose AdaPert, which counters mean collapse by extracting a sparse, perturbation-specific subgraph via differentiable node selection, then suppressing spurious variation in non-responsive genes while emphasizing differentially expressed ones. Across multiple benchmarks, \textsc{AdaPert} outperforms existing baselines, with the largest gains on DEG-aware metrics.

Figures

Figures reproduced from arXiv: 2602.18885 by Chanyoung Park, Hyomin Kim, Jaechang Lim, Junhyeok Jeon, Sang-Yeon Hwang, Seonghwan Kim, Sungsoo Ahn, Woo Youn Kim, Yinhua Piao, Yunhak Oh.

Figure 1
Figure 1. Figure 1: Mean-collapse as a common failure mode in perturba￾tion modeling. For the UQCRB perturbation (n = 114 DEGs), a standard perturbation model shows mean-collapse, where pre￾dicted expression changes shrink toward zero and large effects are underestimated (Left). Our method reduces this bias by using perturbation-specific context and better tracks gene-level expres￾sion changes, especially for strongly express… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of AdaPert (a) The model takes a control cell expression profile x¯c and a perturbation gene p as input. A perturbation￾conditioned subgraph Gcontext is extracted from a biological knowledge graph template G, producing a context representation zp that is combined with the encoded control state zc to predict the perturbed expression profile xˆpert via encoder ENCθ and decoder DECϕ. (b) The adaptive… view at source ↗
Figure 3
Figure 3. Figure 3: Perturbation-Conditioned Subgraph Extraction. Given a perturbed gene p (e.g., UQCRB), the module extracts a perturbation-specific subgraph from the knowledge graph tem￾plate G. First, a textual description of the perturbed gene is retrieved from NCBI and encoded using a language model to obtain a semantic embedding sp. Each node v in the graph is represented by a structural embedding hv computed via messag… view at source ↗
Figure 6
Figure 6. Figure 6: Pathway enrichment analysis of predicted HIRA knock￾down effects.The predicted and ground truth enrichment scores show significant correlation (Pearson r = 0.53, P < 0.001). 5.5. Effect-Size–Dependent Behavior of the Lnon We analyze the interaction between Lnon and perturbation effect size by varying the weight λnon of the non-DEG loss across different perturbation groups ( [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 5
Figure 5. Figure 5: Effect-Size–Dependent Behavior of Lnon. Sensitivity of model performance to the weight λnon across small, medium, and large perturbations, evaluated using pearson∆ and DES@50. −2.0 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 2.0 Normalized enrichment score Myc Targets V1 E2F Targets Myc Targets V2 Mitotic Spindle G2-M Checkpoint Kras Signaling Up Dna Repair Unfolded Protein Response Il-2/Stat5 Signaling Mtorc1 Signalin… view at source ↗
Figure 7
Figure 7. Figure 7: Expression distribution of overall genes and differentially expressed genes in K562 and RPE1 cells. Comparison of gene expression distributions between overall genes and differentially expressed genes (DEGs) in two cell lines. (A-B) K562 cells show right-skewed overall expression (mean=0.54, median=0.37) with 3,015 DEGs exhibiting slightly higher expression (mean=0.64, median=0.43). (C-D) RPE1 cells displa… view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of differentially expressed genes (DEGs) across perturbation effect size categories. (Top row) Stacked bar plots showing the mean number of DEGs (red) and non-DEGs (blue) for each effect size category in K562 (left) and RPE1 (right) datasets. Numbers indicate the mean gene count per category. (Bottom row) Histograms showing the distribution of DEG counts across individual perturbations, colore… view at source ↗
Figure 9
Figure 9. Figure 9: , a large fraction of DEGs lie close to the perturbed gene in the graph. This supports the use of local graph context for perturbation modeling. (a) Top-10 predicted genes (b) Top-20 predicted genes [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Correlation of pathway enrichment between predicted and ground truth responses for HIRA knockdown. Each point represents one of 44 Hallmark pathways, with the x-axis showing ground truth NES and the y-axis showing predicted NES. Colors indicate significance status (FDR < 0.25): gray, non-significant in both; blue, significant in ground truth only; coral, significant in predictions only; red, significant i… 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

23 extracted references · 3 linked inside Pith

  1. [1]

    K., Gautam, D., Bevilacqua, B., Imran, A., Shah, R., Naghipourfar, M., Teyssier, N., Ilango, R., Nagaraj, S., Dong, M., et al

    Adduri, A. K., Gautam, D., Bevilacqua, B., Imran, A., Shah, R., Naghipourfar, M., Teyssier, N., Ilango, R., Nagaraj, S., Dong, M., et al. Predicting cellular responses to perturbation across diverse contexts with state.BioRxiv, pp. 2025–06,

    2025

  2. [3]

    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,

    2024

  3. [5]

    M., Torkar, M., Li, D., and Karaletsos, T

    Istrate, A.-M., Milletari, F., Castrotorres, F., Tomczak, J. M., Torkar, M., Li, D., and Karaletsos, T. rbio1-training scientific reasoning llms with biological world models as soft verifiers.bioRxiv, pp. 2025–08,

    2025

  4. [9]

    D., Simmonds, S

    Pearce, J. D., Simmonds, S. E., Mahmoudabadi, G., Krish- nan, L., Palla, G., Istrate, A.-M., Tarashansky, A., Nelson, B., Valenzuela, O., Li, D., et al. A cross-species genera- tive cell atlas across 1.5 billion years of evolution: The transcriptformer single-cell model.bioRxiv, pp. 2025–04,

    2025

  5. [10]

    S., Quake, S

    Rosen, Y ., Roohani, Y ., Agarwal, A., Samotorˇcan, L., Con- sortium, T. S., Quake, S. R., and Leskovec, J. Universal cell embeddings: A foundation model for cell biology. bioRxiv, pp. 2023–11,

    2023

  6. [12]

    E., Huang, Q., Fang, T., et al

    Szklarczyk, D., Nastou, K., Koutrouli, M., Kirsch, R., Mehryary, F., Hachilif, R., Hu, D., Peluso, M. E., Huang, Q., Fang, T., et al. The string database in 2025: protein networks with directionality of regulation.Nucleic Acids Research, 53(D1):D730–D737,

    2025

  7. [14]

    P., Ektefaie, Y ., et al

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

    2024

  8. [16]

    Data Statistics A.1

    10 Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction A. Data Statistics A.1. Single-Cell Genetic Perturbation Data Predicting how cells respond to genetic perturbations is a key problem in functional genomics (Shalem et al., 2015). It supports many downstream tasks, such as understanding gene function, anal...

    2015

  9. [17]

    Gene expression profiles are measured under single-gene perturbations with matched control cells

    datasets generated using Perturb-seq experiments (Dixit et al., 2016; Replogle et al., 2022; Lotfollahi et al., 2023). Gene expression profiles are measured under single-gene perturbations with matched control cells. Differential expression relative to controls is used as the prediction target. We evaluate on theK562.ReplogleandRPE1.Replogledatasets (Repl...

    2016

  10. [18]

    Perturbations are grouped into small-, medium-, and large-effect categories based on the fraction of differentially expressed genes identified at a significance threshold of p <0.05. This stratification reflects substantial heterogeneity in perturbation responses and enables a more fine-grained evaluation under both sparse and strong transcriptional effec...

    2022

  11. [20]

    The embeddings are available in two variants: Ada (1,536-dim) and Model 3 (3,072-dim), covering 93,800 and 133,736 genes respectively

    embeddings derived from NCBI and UniProt gene descriptions encoded via OpenAI’s text embedding models (Dunefsky et al., 2024). The embeddings are available in two variants: Ada (1,536-dim) and Model 3 (3,072-dim), covering 93,800 and 133,736 genes respectively. Coverage for our datasets is high: 95.8% for K562 and 98.6% for RPE1 HVGs. Table 7.Statistics a...

    2024

  12. [22]

    While effective at capturing global expression shifts, these methods are not explicitly designed to recover sparse gene-level effects

    aim to align control and perturbed cell populations at the distribution level. While effective at capturing global expression shifts, these methods are not explicitly designed to recover sparse gene-level effects. More recently, large-scalefoundation modelshave been introduced for single-cell biology, including (Cui et al., 2024; Hao et al., 2024; Theodor...

    2024

  13. [23]

    Recent studies further explore the integration oftextual and semantic biological knowledge

    leverage gene–gene interaction networks or pathway graphs to propagate perturbation signals through known biological relationships, improving generalization to unseen perturbations. Recent studies further explore the integration oftextual and semantic biological knowledge. Approaches including (Chen & Zou, 2024; Istrate et al., 2024; Wu et al.; Istrate et al.,

    2024

  14. [228]

    exhibit substantially higher expression (mean = 1.34) and a more symmetric distribution. Red and green dashed lines indicate the mean and median, respectively.(2) Perturbation effect size and DEG distributions.To describe variation in perturbation responses, we group perturbations into three categories: small, medium, and large. Effect size is defined as ...

    2000

  15. [1992]

    scgenept: Is lan- guage all you need for modeling single-cell perturbations? bioRxiv, pp

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

    2024

  16. [2009]

    M., Nassar, M., Osi´nski, B., Eksi, R., Yan, Z., Stark, R., Zhang, K., and Grae- pel, T

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

    arXiv

  17. [2015]

    L., Fang, T., Doncheva, N

    Szklarczyk, D., Kirsch, R., Koutrouli, M., Nastou, K., Mehryary, F., Hachilif, R., Gable, A. L., Fang, T., Doncheva, N. T., Pyysalo, S., et al. The string database in 2023: protein–protein association networks and func- tional enrichment analyses for any sequenced genome of interest.Nucleic acids research, 51(D1):D638–D646,

    2023

  18. [2017]

    A systematic comparison 9 Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction of single-cell perturbation response prediction models

    Li, L., You, Y ., Fu, Y ., Liao, W., Fan, X., Lu, S., Cao, Y ., Li, B., Ren, W., Kong, J., et al. A systematic comparison 9 Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction of single-cell perturbation response prediction models. bioRxiv, pp. 2024–12,

    2024

  19. [2018]

    T., Bendidi, I., Russell, C., Hodgson, L., Mesbahi, Y

    Wenkel, F., Tu, W., Masschelein, C., Shirzad, H., Eastwood, C., Whitfield, S. T., Bendidi, I., Russell, C., Hodgson, L., Mesbahi, Y . E., et al. Txpert: Leveraging biochemical relationships for out-of-distribution transcriptomic per- turbation prediction.arXiv preprint arXiv:2505.14919,

    Pith/arXiv arXiv

  20. [2019]

    Subsequent methods, including (Lotfollahi et al., 2023; Adduri et al., 2025), extend this paradigm by conditioning latent variables on perturbation identities and cellular contexts

    learn latent representations of gene expression and infer perturbation effects through shifts in latent space. Subsequent methods, including (Lotfollahi et al., 2023; Adduri et al., 2025), extend this paradigm by conditioning latent variables on perturbation identities and cellular contexts. Related to these approaches, several models formulate perturbati...

    2023

  21. [2023]

    M., Miller, H

    Mejia, G. M., Miller, H. E., Leblanc, F. J., Wang, B., Swain, B., and Camillo, L. P. d. L. Diversity by design: Ad- dressing mode collapse improves scrna-seq perturbation modeling on well-calibrated metrics.arXiv preprint arXiv:2506.22641,

    Pith/arXiv arXiv

  22. [2024]

    and Zou, J

    Chen, Y . and Zou, J. Genept: a simple but effective founda- tion model for genes and cells built from chatgpt.bioRxiv, pp. 2023–10,

    2023

  23. [2025]

    Categorical repa- rameterization with Gumbel-Softmax.arXiv preprint arXiv:1611.01144,

    Jang, E., Gu, S., and Poole, B. Categorical repa- rameterization with Gumbel-Softmax.arXiv preprint arXiv:1611.01144,

    Pith/arXiv arXiv