Predicting transcriptome responses to unseen perturbations in unseen cell contexts (virtual cell models)

• Linear baselines (Nature Methods 2025). Ahlmann-Eltze, Huber and Anders compared five single-cell foundation models (including scGPT and scFoundation) and two other deep-learning models (GEARS, CPA) with deliberately simple baselines for single and double genetic perturbations and found none outperformed the baselines; for unseen genes a linear model performed about as well as the deep models. A pre-review noted that different linear models were used for single and double perturbations, so details of the comparison are debated.
• Systematic variation (Systema, Nature Biotechnology 2025). Simple baselines perform comparably to existing methods because of systematic differences between perturbed and control cells; predicting responses to unseen perturbations is substantially harder than standard metrics suggest.
• PertEval-scFM (preprint, 2024). Single-cell foundation model embeddings give no consistent improvement over baselines, especially under distribution shift, and all models struggle with strong or atypical perturbation effects.
• Arc Virtual Cell Challenge 2025. About 300,000 H1 human embryonic stem cells with 300 CRISPRi perturbations; models trained on other cell types had to predict held-out responses. Participation was reported in the thousands of registrants and hundreds of final submissions (organiser figures). The winning and prize-winning entries are reported (in challenge announcements and press coverage, not a refereed analysis) to have combined deep learning with protein embeddings, public perturbation data and statistical features; team names and prize categories here follow those announcements and were not verified against a refereed report. Reports of the challenge state that no model consistently beat naive baselines on every metric; a statistics preprint (Liu et al., arXiv:2511.16954) shows the perturbation discrimination score is highly sensitive to the distance measure and scaling, so metric design is itself open.
• Arc Virtual Cell Challenge 2026. Opened in August 2026 with a zero-shot design (no training set released; new data in multiple cell lines); no results were available as of October 2026.
• Not shown. No model satisfies the target above under baseline-controlled metrics.

A reliable virtual cell would let researchers triage CRISPR and drug screens in silico, predict combination effects, prioritise therapeutic targets and anticipate off-target transcriptional effects, reducing experimental cost. Today's models are not reliable enough for these uses, and the main current value of the problem is establishing honest benchmarks.

A pre-registered, prospective evaluation (for example a later Arc challenge or an independent benchmark with newly generated data) in which one model meets criteria (i)-(iv) on at least three held-out screens, reported with the baselines above and with a metric analysis showing the result is not an artefact of systematic variation or distance-metric choice, then confirmed by a refereed publication.