Perturbation-response prediction methods: a map of designs recorded in the library

Research question

• Which families of perturbation-response prediction methods are recorded in the library, and which are peer reviewed versus preprints?
• Where should readers look for evidence on whether these methods outperform simple baselines?

Current understanding

Our reading of the library records: perturbation-response prediction methods span variational autoencoders, compositional autoencoders, graph neural networks, optimal transport, disentanglement, knowledge-graph, diffusion and flow-matching designs, with the more recent and larger models mostly recorded as preprints. This page maps designs only; it does not rank methods, and benchmark papers in the library report that simple baselines are hard to beat under some evaluations. Certainty: high that the records describe the designs as stated; no performance comparison is made.

Scope and how to read this page. Scope: library records only (title, venue, one-line summary and verification source); we have not read the full texts for this page, and every statement below is limited to what those records state. Preprints are marked and are not peer reviewed. This page makes no claim about glioma or other tumours. This is a map of designs, not a leaderboard. VC0013 VC0048

Earlier peer-reviewed designs. scGen is recorded as an early VAE-based single-cell perturbation-response predictor (Nature Methods 2019). CPA is recorded as a compositional perturbation autoencoder for drug, dose and combination responses (Molecular Systems Biology 2023). GEARS is recorded as a knowledge-graph graph neural network for multigene perturbation outcomes (Nature Biotechnology 2023). CellOT is recorded as neural optimal transport for single-cell perturbation responses (Nature Methods 2023); its verification note records that an erratum was published and that its impact on the cited conclusions should be checked, which we have not done. biolord is recorded as disentangled representations for counterfactual cell prediction (Nature Biotechnology 2024). Certainty: high that the records state this. VC0015 VC0014 VC0013 VC0016 VC0017

More recent peer-reviewed designs. TxPert uses multiple knowledge graphs for transcriptomic perturbation prediction (Nature Biotechnology 2026). State (Arc Institute) is recorded as a state-embedding plus transition model across 70 contexts (Cell 2026); the earlier bioRxiv version is a separate record (preprint). UniPert-G2CP is recorded as unified genetic and chemical perturbagen embeddings with CRISPR-to-drug phenotype transfer (Cell 2026; month-level date only). Certainty: high that the records state this. VC0018 VC0019 VC0020 · preprint VC0130

Preprint designs (not peer reviewed). X-Cell is recorded as a 4.9B-parameter diffusion model with perturbation scaling laws. PerturbLDM is recorded as a latent diffusion model pretrained on Tahoe-100M for conditional response generation across drug, dose and cell line. AlphaCell is recorded as a genome-wide virtual-cell world model using optimal-transport conditional flow matching. AetherCell is recorded as unifying clinical RNA-seq and perturbation assays for generative response prediction and in vivo repurposing. VCHarness is recorded as an autonomous AI coding agent that searches and refines perturbation-response model architectures. Certainty: moderate; self-described in preprints, no independent evaluation recorded in the library. VC0033 · preprint VC0138 · preprint VC0137 · preprint VC0136 · preprint VC0135 · preprint

Where to look for baseline comparisons. A Nature Methods paper reports that deep-learning gene perturbation effect prediction does not yet outperform simple linear baselines (VC0048), and a Genome Biology comparison of expression-forecasting methods reports that it is uncommon for methods to beat simple baselines (VC0145). A Nature Biotechnology paper takes the counter-position that deep models can outperform baselines on calibrated metrics (VC0146). These studies differ in datasets and metrics; our commentary on how they differ is on the linear-baselines page. Certainty: high that the papers state this. VC0048 VC0145 VC0146

Evidence map

Primary studies & datasets

• VC0015 scGen predicts single-cell perturbation responses. — Lotfollahi M, Wolf FA, Theis FJ., Nature methods, 2019
• VC0014 Predicting cellular responses to complex perturbations in high-throughput screens. — Lotfollahi M et al., Molecular systems biology, 2023
• VC0013 Predicting transcriptional outcomes of novel multigene perturbations with GEARS. — Roohani Y, Huang K, Leskovec J., Nature biotechnology, 2023
• VC0016 Learning single-cell perturbation responses using neural optimal transport. — Bunne C, Stark SG, Gut G, Del Castillo JS, Levesque M, Lehmann KV, Pelkmans L, Krause A, Rätsch G., Nature methods, 2023
• VC0017 Disentanglement of single-cell data with biolord. — Piran Z, Cohen N, Hoshen Y, Nitzan M., Nature biotechnology, 2024
• VC0018 TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects. — Wenkel F et al., Nature biotechnology, 2026
• VC0019 Predicting cellular responses to perturbation across diverse contexts with State. — Adduri AK et al., Cell, 2026
• VC0130 UniPert-G2CP bridges genetic and chemical screens from molecular representation to phenotype modeling — Li et al., Cell, 2026

Related models & tools

• VC0020 · preprint Predicting cellular responses to perturbation across diverse contexts with State — Adduri AK et al., bioRxiv, 2025
• VC0033 · preprint X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models — Wang C et al., bioRxiv, 2026
• VC0137 · preprint Towards building a World Model to simulate perturbation-induced cellular dynamics by AlphaCell — Chuai et al., bioRxiv, 2026
• VC0138 · preprint PerturbLDM: conditional latent diffusion for modelling single-cell perturbation responses — Yu et al., bioRxiv, 2026
• VC0136 · preprint AetherCell: A generative engine for virtual cell perturbation and in vivo drug discovery — Li et al., bioRxiv, 2026
• VC0135 · preprint Harnessing AI to Build Virtual Cells (VCHarness) — Cheng et al., bioRxiv, 2026

Perturbation & validation evidence

• VC0048 Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. — Ahlmann-Eltze C, Huber W, Anders S., Nature methods, 2025
• VC0145 A comparison of computational methods for expression forecasting. — Kernfeld E, Yang Y, Weinstock JS, Battle A, Cahan P., Genome biology, 2025

19 library items cited on this page; 7 are preprints (not peer reviewed).

Related reviews

• Glioblastoma cell states and AI virtual cells
• GBM cell-state atlases and the virtual-cell foundation (draft, pending review)

Entities on this page

• AlphaCell Models
• CellOT Models
• CPA Models
• GEARS Models
• Simple linear baseline Methods & concepts
• Perturbation response prediction Methods & concepts
• PerturbLDM Models
• scGen Models
• State Models
• Tahoe-100M Datasets
• TxPert Models
• Virtual cell world model Methods & concepts
• X-Cell Models

What remains uncertain

The library records do not give a like-for-like comparison of these methods on one dataset and metric, so no ordering of methods can be read from this page. Benchmark papers disagree about whether deep models outperform simple baselines. VC0048 VC0146

The records here mostly do not state the cellular contexts of each method's training data. The two large perturbation datasets recorded in the library are cell-line based (K562/RPE1 for VC0040; 50 cancer cell lines for VC0041), so how these methods behave in other cellular contexts is not established by the records listed here. VC0040 VC0041 · preprint

Related questions

• When papers report that a mean or linear baseline matches deep models, are they comparing the same object, with the same metric, on the same kind of perturbations?
• Which mechanisms could make a perturbation-agnostic predictor look good (shared response, noise, target gene, metric), and which papers actually test which?
• How are virtual cell models validated?
• Do deep learning models beat linear baselines for perturbation prediction?
• Which large perturbation datasets and profiling platforms are in the library, and what do their records state about scale and cellular system?
• Which statements are dataset-producer descriptions rather than independent evaluations?
• What is an AI virtual cell model?
• Which virtual cell models have been applied to glioblastoma?