AI Virtual Cell

The "AI Virtual Cell" (AIVC) — a learned model of a cell that predicts how perturbations change cell state — moved in 2025–2026 from a concept-paper goal into a funded, benchmarked, multi-modality landscape.

What the virtual cell is

The AI Virtual Cell is an ML system that, given a cell state and a perturbation (a drug, a genetic edit), predicts the resulting state. The framing was crystallized by a 2024 Cell perspective ("How to build the virtual cell with artificial intelligence") [1]. The value thesis is the same as the rest of AI drug discovery: move failure out of the wet lab and into simulation, collapsing the 10+ year / billion-dollar discovery cycle. See ai-for-science for the cross-domain framing and aifs-biology for the skeptical baseline (single-cell foundation models repeatedly failing to beat simple methods).

The modality split — and why it matters

The field has converged on five distinct output modalities, and the choice is not cosmetic — it changes what the model can be trusted for [local: virtual_cell_video_comparison.md]:

Why image often wins in practice. Image (Cell Painting) is far cheaper (~$0.5–1/well vs ~$6–10/well for RNA-seq) and far more reproducible (57–83% vs 16–35% for transcriptome), and morphology is a direct readout of function rather than a proxy that must be re-inferred [local]. Why transcriptome still matters. It reveals mechanism — which pathway activated — that morphology cannot. Video adds time (fate decisions, drug-response dynamics) but is the scarcest and hardest-to-validate mode.

The honest structural finding

A 2026 finding that tempers the enthusiasm: cross-modal prediction accuracy does not imply a clean gene→phenotype mapping. Single-gene correlations with morphology are generally weak (r < 0.2); morphology emerges as a **polygenic, distributed** "summary statistic" integrating many weak transcriptional signals [local]. Multimodal fusion is where the real gains are: combining modalities predicts ~59% of assays (AUROC > 0.7) versus ~7–8% for any single modality [local].

The 2026 landscape

International (funding figures are directional, from local research):

• Xaira Therapeutics — $1B+; X-Cell (~4.9B params), a scaling-law-driven virtual cell.
• Arc Institute — $650M+; STATE / Stack, context-learning single-cell foundation models.
• Cellular Intelligence — ~$57M; universal virtual-cell signal model.
• Noetik — ~$50M; OCTOVC, a spatial virtual cell.
• Tahoe Therapeutics — Tahoe-100M, a 100M-cell perturbation dataset.
• genbio-ai — AIDO, a multi-scale "digital organism" spanning molecule → cell → tissue (AIDO Cell, Aug 2026).

China: 百曜科技 (AURA CellOS, ~12B params, JEPA), 无界进化 (OCOO-T), 华源智因, 深度细胞, plus big-lab efforts (阿里达摩院 灵枢 cell model; 腾讯 UniPert-G2CP).

recursion-pharmaceuticals occupies the adjacent data-first pole — industrialized phenomics imaging rather than a single virtual-cell model.

Where this is going

1. Accountability arrives. The zero-shot 2026 Virtual Cell Challenge and MVCBench mean claims will be judged on public, baseline-anchored benchmarks — a shift from self-reported leaderboards [local].
2. Multi-scale is the new bet. genbio-ai's AIDO and similar efforts move from single-layer models toward integrated molecule→cell→tissue systems — the "virtual organism" ambition.
3. The baseline problem persists. As aifs-biology documents, "foundation model" branding has repeatedly failed to beat well-tuned baselines in single-cell settings; the virtual cell inherits that burden until zero-shot benchmarks say otherwise.

Sources

• [1] Bunne et al., "How to build the virtual cell with artificial intelligence," Cell (2024).
• [local] virtual_cell_video_comparison.md — modality/company/evaluation landscape (compiled 2026-08-24)
• [local] 2026-08-24-summary.md — daily research summary