henhao
Sign in
Systems BiologyCell
Peer-reviewed journal article

Predicting cellular responses to perturbation across diverse contexts with STATE

Abhinav AdduriiD, Dhruv Gautam, Beatrice BevilacquaiD, Mohsen NaghipourfariD, Alishba Imran, Rohan Shah, Noam TeyssieriD, Rishi VermaiD, Christopher Carpenter, Basak EraslaniD, Francis Chalissery, Rajesh Ilango, Vishvak SubramanyamiD, Chiara Ricci-TamiD, Sanjay Nagaraj, Aidan F. WintersiD, Mingze DongiD, Stefanie Fellinger, Adam Krejčí, Tilmann BürckstümmeriD, Sravya Tirukkovular, Jeremy Sullivan, Brian S. PloskyiD, Nicholas D. Youngblut, Jure LeskoveciD, , Yusuf RoohaniiD

Abstract

Cellular responses to perturbations are a cornerstone for understanding biological mechanisms and selecting drug targets. While machine learning models offer tremendous potential for predicting perturbation effects, they currently struggle to generalize to unobserved cellular contexts. Here, we introduce SO_SCPLOWTATEC_SCPLOW, a transformer model that predicts perturbation effects while accounting for cellular heterogeneity within and across experiments. SO_SCPLOWTATEC_SCPLOW predicts perturbation effects across sets of cells and is trained using gene expression data from over 100 million perturbed cells. SO_SCPLOWTATEC_SCPLOW improved discrimination of effects on large datasets by more than 30% and identified differentially expressed genes across genetic, signaling and chemical perturbations with significantly improved accuracy. Using its cell embedding trained on observational data from 167 million cells, SO_SCPLOWTATEC_SCPLOW identified strong perturbations in novel cellular contexts where no perturbations were observed during training. We further introduce Cell-Eval, a comprehensive evaluation framework that highlights SO_SCPLOWTATEC_SCPLOWs ability to detect cell type-specific perturbation responses, such as cell survival. Overall, the performance and flexibility of SO_SCPLOWTATEC_SCPLOW sets the stage for scaling the development of virtual cell models.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
100
FWCI
161.45
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (10 Jul 2025) · CC BY 4.0

State, a set-based transformer trained on over 100 million perturbed cells, predicts perturbation responses in new cellular contexts better than mean, linear and deep learning baselines.

Deep learning perturbation models have not reliably beaten linear baselines when transferring effects to unseen cell contexts, partly because within-population heterogeneity and cross-dataset technical noise mask true effects. State addresses both by attending over covariate-matched cell sets and by using a cell embedding learned from observational data to absorb technical variation.

Key results

  1. ShownWith 30% of target-context perturbations visible in training, State ranked perturbations and predicted expression shifts in held-out cell lines or donors better than baselines. The largest margins were on chemical and cytokine data.

    Perturbation discrimination: absolute gains of 54% (Tahoe-100M) and 29% (Parse-PBMC); matched the perturbation mean on Replogle-Nadig. Pearson delta correlation: 63% above baselines on Tahoe-100M, 47% on Parse-PBMC, 5% on Replogle-Nadig. Comparators: context and perturbation means, linear, CPA, scVI, scGPT. · 2.2. State outperforms baselines in predicting…; Perturbation Discrimination Score; Pearson Delta Correlation

  2. ShownState identified differentially expressed (DE) genes and ranked perturbations by overall effect size better than baselines. On genetic data, individual baselines still won specific metrics.

    DE-gene precision-recall AUC 184% above the next best method on Replogle-Nadig; top-DE overlap twice scGPT on Tahoe-100M and 43% above Linear on Parse-PBMC; effect-size Spearman 70% higher on Tahoe-100M (near 0.8), 53% on Parse-PBMC, 22% on Replogle-Nadig. On Replogle-Nadig, Linear beat State on DE overlap by 20% and context mean on fold change by 12%. · 2.2. State outperforms baselines in predicting…; DE Overlap Accuracy; PR-AUC; Effect sizes

  3. ShownSE embeddings separated perturbation labels better than other single-cell foundation model embeddings and raw counts. After Tahoe-100M pretraining and fine-tuning, ST with SE inputs ranked zero-shot perturbation effect sizes in unperturbed contexts better than ST on expression and the perturbation mean.

    Extrinsic perturbation classification: over 6% absolute gain vs scFoundation, scGPT, UCE, Transcriptformer and raw counts (two held-out sets). Zero-shot (2,102 perturbations, 5 datasets, 40 contexts): ST+SE beat ST+HVG (expression input) and perturbation mean; over 17% gain, Spearman above 0.5, on Parse-PBMC and Replogle-Nadig; mean 15% gain over five metrics on Parse-PBMC. · 2.3. State embeddings enhance zero-shot…; Intrinsic Evaluation; Extrinsic Evaluation

  4. ShownOn five held-out Tahoe-100M cell lines, State captured cell line-specific drug responses, drug similarity structure and cell survival better than the mean baselines.

    Trametinib, top 100 DE genes per line: State recovered over 30% vs fewer than 7% for both mean baselines (about 20 extra line-specific genes). Drug-similarity clustering Adjusted Rand Index over 0.7 vs less than 0.1 for perturbation mean. SVM survival prediction: Pearson 0.52 vs 0.2 (context mean) and 0.31 (perturbation mean). · 2.4. State can detect cell type-specific…; Drug Similarity Analysis; Survival Prediction Analysis

  5. ShownSet-level self-attention mattered: larger cell sets lowered validation loss up to an optimum, and removing attention or replacing it with mean pooling degraded performance.

    Tahoe-100M ablation with set size times batch size fixed at 16,384: held-out validation loss improved as set size grew to an optimum of 256. Full ST beat a mean-pooling pseudobulk variant, a set-size-1 variant and a variant without self-attention (loss values not given in text). · 2.1. Building the State Transition model for…; Cell Set Scaling; Figure legends

Methods, data and reproducibility

State couples a State Transition (ST) transformer, which attends across covariate-matched control cell sets, is conditioned on perturbation and is trained with a maximum mean discrepancy loss, with a 600M-parameter State Embedding (SE) model pretrained on 167 million observational cells. ST's solution family asymptotically contains the quadratic-cost optimal transport map.

  • Tahoe-100M drug perturbation single-cell data from cancer cell lines; five lines (C32, HOP62, HepG2/C3A, Hs 766T, PANC-1) held out, plus leave-one-tissue-out splits · human (cancer cell lines) · n = 50 cell lines; 380 drugs; 1,138 conditions
  • Parse Biosciences cytokine signaling perturbations of peripheral blood mononuclear cells (Parse-PBMC); 4 donors held out · human (donor PBMCs; human gene set) · n = 12 donors; 18 cell types; 90 cytokines; 10 million cells
  • Replogle et al. (2022) and Nadig et al. (2025) genome-scale genetic knockdown Perturb-seq, filtered for on-target knockdown; leave-one-cell-line-out · human (cell lines incl. K-562 and Jurkat) · n = 2,024 perturbations; four cell lines
  • Smaller perturbation datasets used in ST experiments (Jiang et al. and McFaline et al. genetic screens; Srivatsan), including zero-shot query sets · human (human protein-coding gene set)
  • Observational corpus for SE pretraining: Arc scBaseCount, CZ CELLxGENE and 20 non-test Tahoe-100M cell lines; dataset-level validation split · human · n = 167 million cells; 14,420 AnnData files
  • Tahoe-100M per-well cell survival (relative fitness vs plate-matched DMSO) for held-out lines, used to train a survival regressor · human (cancer cell lines) · n = 35 drugs (all 3 doses) held out per repeat; 10 repeats
Methods & tools
State Transition (ST): bidirectional LLaMA or GPT2 transformer over cell sets, State Embedding (SE): 600M-parameter transformer encoder with MLP decoder, Cell-Eval evaluation framework, Cell-Load (filter_on_target_knockdown), ESM-2 esm2_t48_15B_UR50D gene embeddings, Scanpy preprocessing (normalize_total, log1p, highly_variable_genes, pca), geomloss energy-distance MMD loss, PyTorch Lightning with DDP, AMP and Flash Attention 2, Baselines: linear ridge model, CPA, scVI, scGPT, Embedding baselines: scFoundation, UCE, Transcriptformer, Wilcoxon rank-sum DE with Benjamini-Hochberg FDR, scikit-learn support vector regressor (rbf kernel)
Code
https://github.com/ArcInstitute/Statehttps://github.com/ArcInstitute/State-reproducehttps://github.com/ArcInstitute/Cell-evalhttps://github.com/ArcInstitute/Cell-load
Data & accessions
https://huggingface.co/arcinstitute/SE-600Mhttps://huggingface.co/arcinstitute/ST-Tahoehttps://huggingface.co/arcinstitute/ST-Parse
Software & environment
PyTorch Lightning with DDP and automatic mixed precision; LLaMA or GPT2 backbones from HuggingFace; Flash Attention 2 with bf16 for SE; Scanpy, geomloss, scikit-learn; no package versions or container specified.
Assay conditions & compute
SE trained for 4 epochs on 4 nodes of 8 NVIDIA H100 GPUs each; effective batch 3,072 (per-device 24, gradient accumulation 4 steps). ST training hardware and wall-clock times not reported.
Reproducibility gaps
The data URLs are released model weights (SE-600M, ST-Tahoe, ST-Parse); no links or accessions for the perturbation or SE pretraining datasets are printed; ST hyperparameters sit in Tables 3 and 4; versions and runtimes unstated.
Limitations (6)
  • Usability on wholly unseen datasets, where none of the cell contexts appeared in training, was not tested; the zero-shot setting holds out one context inside a query dataset after fine-tuning on its other contexts. · stated by the authors
  • Subtle-effect metrics, such as how accurately individual DE genes are called, are more affected by dataset size and quality; in small genetic sets (McFaline, Jiang) baseline performance was near zero and gains on other metrics varied with data quality. · stated by the authors
  • On genome-scale genetic screens from few cell lines, State's margin over mean baselines was small: it only matched the perturbation mean in discrimination, and the linear model and context mean beat it on DE overlap (20%) and fold change (12%). · stated by the authors
  • The optimal transport result is asymptotic (infinite set size) and assumes bounded densities on convex compact supports; the unconstrained solution family holds infinitely many other maps, so OT is not guaranteed to be learned and implicit bias is uncharacterized. · stated by the authors
  • Trametinib dose is inconsistent (0.5 µM in Results vs 0.05 µM in figure legends), and this case was picked among perturbations ranked by gain over baselines, so it illustrates rather than estimates typical cell type-specific accuracy. · noted in this summary
  • Perturbations are one-hot encoded, so tests only cover perturbations seen in other contexts; many gains are given as percentages relative to the best baseline, with few absolute values or intervals in text, making absolute accuracy hard to judge. · noted in this summary

Who should read:Read this if you build or evaluate single-cell perturbation models and need a set-based transformer, a perturbation-aware cell embedding, and Cell-Eval metrics for transfer to new cell contexts.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLu '26Li '25Yu '25Chuai '26Ji '24Palla '26
71% semantic matchBioinformatics
Mingkun Lu, Nanxin You, Hongning Zhang+6 more·bioRxiv·Apr 28, 2026