A validation ladder for virtual-cell claims: a proposed way to label how far a claim has been tested

Research question

• How far has a virtual-cell claim been tested: against held-out computational data, in a wet-lab assay, in animals, in patient data, or prospectively?
• Which library records show which kind of test, and where does the library say nothing?

Current understanding

Our proposal: label each virtual-cell claim with the highest level at which it has actually been tested, from V0 (computational only) to V5 (prospective clinical evidence), and never infer a higher level from a lower one. Most benchmark and challenge records in the library sit at the held-out-data level (V1); one preprint record calls a model-nominated combination validated without recording the assay (VC0139). We do not assign records to V3–V5 here because that needs full-text reading we have not done. Certainty: the ladder is our definition; record-level placements are provisional.

Scope and status. This page is a GlioState proposal for how to label validation level. The six levels below are our own editorial definitions, not a published standard and not yet reviewed; they are meant to make the highest level of validation in each record explicit, never to raise it. Placements use library records only (title, summary, verification note). VC0004

The levels (our definitions). V0: computational only, with no held-out or external data. V1: tested against held-out or external computational data, including benchmarks and challenges. V2: tested in an in vitro wet-lab assay. V3: tested in animals. V4: tested against retrospective patient data. V5: prospective clinical evidence. A claim is labelled with its highest level reached, and a V1 result does not imply V2. VC0004

What sits at V1 in the library. The Virtual Cell Challenge is introduced by an Arc Institute perspective as a Turing test for the virtual cell (VC0002), and the 2026 challenge paper benchmarks zero-shot generalisation across cellular contexts (VC0058). Benchmark and critique papers evaluate perturbation-prediction methods on held-out perturbations (VC0048, VC0052). All of these are tests against held-out computational data. The 2025 challenge wrap-up record reports lessons about metric gaming (VC0105), a reminder that a V1 result depends on the metric used. Certainty: high that the records are benchmark or challenge records; this does not say what any specific model's V1 result was. VC0002 VC0058 VC0048 VC0052 VC0105

A possible V2 statement. A preprint (VC0139) records a Transformer model (Mosaic) that nominated a triple-TF combination, which the record describes as validated; the library does not record the assay, so we do not place this at V2 yet. The record is a preprint, not peer reviewed. VC0139 · preprint

Where the library says nothing. A review of AI virtual cell models in preclinical research discusses technical pathways and validation (VC0004; we have not read how it grades validation). We do not assign any record to V3–V5 on this page, so the absence of records at those levels here is a limit of this page, not a finding about the field. VC0004

Evidence map

Primary studies & datasets

• VC0004 AI-driven virtual cell models in preclinical research: technical pathways, validation mechanisms, and clinical translation potential. — Ma C, Zhang H, Rao Y, Jiang X, Liu B, Sun Z, Song Z, Gao Y, Cui Y, Liu X, Li Z., NPJ digital medicine, 2025
• VC0002 Virtual Cell Challenge: Toward a Turing test for the virtual cell. — Roohani YH et al., Cell, 2025
• VC0058 Virtual Cell Challenge 2026: Benchmarking zero-shot generalization across cellular contexts. — Arc Virtual Cell Initiative Team, Goodarzi H., Cell, 2026
• 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
• VC0052 Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation. — Viñas Torné R, Wiatrak M, Piran Z, Fan S, Jiang L, Teichmann SA, Nitzan M, Brbić M., Nature biotechnology, 2025

Related models & tools

• VC0139 · preprint Engineered regulatory states resolve context-dependent genetics in macular degeneration — Metzner E, Guruvayurappan K, Southard KM, Ardy RC, Dey KK, Norman TM., bioRxiv, 2026
• VC0105 Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections — Arc Institute, Arc news, 2025

7 library items cited on this page; 1 are preprints (not peer reviewed).

Entities on this page

• AI virtual cell Models
• Perturbation response prediction Methods & concepts
• Virtual Cell Challenge Methods & concepts
• Zero-shot evaluation Methods & concepts

What remains uncertain

The ladder is a proposal with our own definitions; whether the six levels and their boundaries are right, and whether V3–V5 should be split by system or tissue, is undecided pending editorial review. VC0004

Record-level placement is provisional: the placements rest on titles, summaries and verification notes, and a full-text reading could move a record up or down. VC0058 VC0139 · preprint

Related questions

• How are virtual cell models validated?
• Do deep learning models beat linear baselines for perturbation prediction?
• What do the cell-eval metrics used by the Challenge actually reward, when checked against code and public data?
• Where do the two main baseline critiques (VC0048, VC0052) agree, conflict, or rest on unstated assumptions?
• Which evaluation pitfalls do the benchmark and critique papers in the library identify?
• For each pitfall, which record is the source and how mature is that source (peer reviewed or preprint)?