Toward Generative Virtual Cells — Lewis & Zueco (AIXC Research)¶
One-line description — A co-evolutionary framework where a virtual-cell world model (predicts experimental outcomes) and a perturbation planner (chooses which experiments to run) jointly adapt on inner and outer timescales, with validation-gated architecture search closing the loop.
- Authors: David Scott Lewis, Enrique Zueco
- Affiliation(s): AIXC Research
- Track / workshop: ICLR 2026 — Gen² workshop (Generative AI in Genomics)
- OpenReview: openreview.net Gen2 ICLR 2026 PDF
- Code: TBD — not released at workshop time
- Status at ICLR 2026: Gen² workshop accepted paper (Apr 27, 2026)
What it does¶
The paper argues that "virtual cell" research treats the world model and the experimental planner as separable, when in practice they are tightly coupled: a planner only proposes informative perturbations if the world model is calibrated, and a world model only improves on perturbations the planner actually selects. The contribution is a co-evolutionary training loop that updates both jointly, with a validation-gated mechanism for safely changing the world-model architecture mid-loop.
How it works¶
Core idea. Treat the virtual-cell world model and the experimental planner as a co-evolutionary system updated on two timescales — fast joint adaptation in response to each experiment, and slower validation-gated changes to the world-model architecture itself.
Inputs / outputs. Inputs are a current world-model checkpoint, a candidate perturbation library (Perturb-seq-style genetic or small-molecule), and a held-out validation set; outputs are (a) selected next perturbations to run, (b) updated world-model weights, and (c) accepted/rejected architecture proposals.
Key innovation. Prior virtual-cell pipelines (scGen, CPA, GEARS, and the broader Arc / Recursion virtual-cell wave) train the world model on fixed perturbation datasets and treat experimental design as a downstream step; this paper argues the two should be jointly adapted, with the planner's exploration strategy conditioned on world-model uncertainty and recent error patterns, and architecture moves gated by a validation criterion to prevent catastrophic regression.
Parameters / training details. Proof-of-concept uses a synthetic Perturb-seq simulator and a small MLP / transformer world model; planner exploration is uncertainty-weighted; the outer-loop architecture-change criterion compares validation loss + capacity constraints before commit (TBD — exact thresholds in workshop PDF).
Canonical experiment. On the toy Perturb-seq simulator, the co-evolutionary loop is compared against a fixed-world-model baseline and a fixed-planner baseline; the paper reports that joint adaptation under validation gating outperforms both, with the architecture-change mechanism preventing the runaway drift that plagues naive joint training (TBD — exact numbers behind workshop PDF).
Headline benchmarks¶
This is a position + proof-of-concept paper, not a benchmark paper. The Gen² accept frames the contribution as a paradigm for "safe model adaptation" coupled with experimental design. Quantitative comparisons against fixed-world-model baselines are reported in the workshop PDF on toy Perturb-seq simulators (TBD — exact numbers behind PDF parsing).
Where it fits in the corpus¶
- AACR 2026 virtual-cells axis: direct match — pair with the AACR virtual-cell session and with any cell-state-engineering pharma demos.
- NeurIPS 2026 LMRL: the natural next venue for the diffusion-based + real-Perturb-seq version this paper telegraphs.
- ICML 2026 GenBio: workshop-level cousin contributions on perturbation planners likely.
- AACR Pancreatic / Brain Tumors 2026: longer-term — perturbation planning over tissue-specific cell atlases is the obvious oncology adapt.
Notes¶
The paper is workshop-stage and explicitly speculative — it is included as the concept anchor for the Gen² "engineering cellular states" theme more than as a method release. The Lewis/Zueco framing — "world model + planner, jointly adapted under validation gating" — is the architecture pattern that AACR-vault agentic-AI dossiers should expect to start emerging from labs like Cradle, Recursion, and Insitro within 6–12 months.
FM comparison & 2026 status¶
Where it sits in the FM landscape. Lewis & Zueco propose a generative virtual-cell model plus joint planner — not a benchmark, not an adapter, and not a pre-trained encoder. It belongs to the small cohort of systems papers that try to close the loop from world model to next experiment, alongside Arc Institute's STATE virtual-cell model release (Arc Institute, 2025) and the Cradle / Recursion in-house planners. Concretely, the world-model side is in the same architectural family as scGPT-perturbation modes, GEARS, and CPA; the planner side is in the active-learning / Bayesian-optimization-over-Perturb-seq tradition.
Direct peers — generative virtual-cell models / perturbation generators.
| System | World model | Planner / loop | Substrate | Release |
|---|---|---|---|---|
| Lewis & Zueco (this paper) | small MLP / transformer + validation-gated architecture search | uncertainty-weighted, jointly adapted | toy Perturb-seq simulator | ICLR 2026 Gen² workshop |
| STATE (Arc Institute) | transformer pretrained on Tahoe-100M + scBaseCount | offline (model only) | 600M-cell Arc Virtual Cell Atlas | Arc Institute, 2025 |
| scGPT generation mode | masked-token autoregressive over scRNA-seq | none — open-loop | CELLxGENE corpus | Cui et al., Nature Methods 2024 |
| GEARS | GNN over gene-perturbation graph | open-loop | Perturb-seq screens | Roohani, Huang, Leskovec Nature Biotech 2023 |
| scFoundation generation modes | 100M-cell pretrained autoregressive | open-loop | scFoundation pretraining mix | Hao et al. 2024 |
| scBaseCamp / scBaseCount | n/a — atlas used as training substrate | AI agent-curated metadata | 500M cells, 21 organisms | Arc Institute scBaseCount, 2025 |
The differentiator vs. STATE (the closest 2025–2026 production-scale analogue) is the joint adaptation under validation gating — STATE is trained offline on a frozen Tahoe-100M snapshot, while Lewis & Zueco argue the planner and world model must be updated together as new experiments come in. This is the paper's load-bearing methodological claim.
What changed in 2025–2026.
- STATE release (Arc Institute, 2025). First production-grade virtual-cell model trained on the Arc Virtual Cell Atlas; Lewis & Zueco's contribution is now legible as "the loop closure STATE doesn't (yet) implement" (Arc Institute STATE announcement).
- Tahoe-100M (Feb 2025). 100M cells × 1,200 drugs × 50 cancer lines — the substrate any future co-evolutionary loop would target (Vevo Tahoe-100M release).
- Virtual Cell Atlas launch (Feb 2025). Combined ≥300M cells (now ≥600M) across Tahoe-100M + scBaseCount, the natural deployment substrate for any planner / world-model loop.
- Linear-baseline reckoning (Ahlmann-Eltze & Huber, Nature Methods 2025). Casts a long shadow on any 2026 generative-virtual-cell paper that does not include the mean-of-training-perturbations baseline as a control (Nature Methods article).
- No retraction events, but the field-wide critique narrows what "outperforms baseline" can mean — Lewis & Zueco's toy-simulator result must be read against this caveat.
Cross-AACR relevance. Direct match to the AACR 2026 ED03 virtual-cells session, particularly the perturbation-prediction-and-in-silico-screens family (6 posters including #1464 CELLama-Perturb and #4181 Turbine Virtual Lab) and the digital-twins family (3 posters). The AACR scGPT dossier and Geneformer dossier are the comparator world-model entries. The "planner" framing maps directly onto AACR's AT02 agentic-AI session, which discussed AI-as-experimentalist autonomy — Lewis & Zueco's validation-gated architecture move is the methods-side answer to AT02's open question about how agents should propose their own next experiments without runaway drift.