Lab Newsletter — July 7, 2026: Big Money Meets the Hard Tests

Two forces are pulling on AI-for-biology at once: capital is arriving at industrial scale, and the honest tests are getting harder to game. Today’s items sit on that fault line.

🏭 Industry goes all-in on agentic labs

NVIDIA and Eli Lilly are standing up a co-innovation lab — up to $1B over five years — to build next-generation biology and chemistry foundation models on NVIDIA BioNeMo, wiring Lilly’s agentic wet labs to computational dry labs for round-the-clock, “scientist-in-the-loop” experimentation, robotics and physical AI included. And it isn’t a one-off: NVIDIA and Thermo Fisher are pushing to make instruments themselves intelligent with multi-agent workflows that generate protocols and run experiments, while a wave of YC startups is building “agentic drug companies” on top of biological foundation models. Why it matters for the lab: the agent-first, self-driving-lab thesis we’ve been building — REEF, Agent-Lens, BioEngine, Hypha — is now exactly where the industry is pouring capital. That’s validation, and a reminder that our edge is openness and closing the loop on real cells. (NVIDIA+Lilly was unveiled in January; it’s part of a 2026-long industrial turn.)

🧫 Virtual cells keep scaling — and the baselines keep them honest

The perturbation-prediction field keeps shipping models, with Arc’s STATE (trained on ~170M observational and 100M+ perturbational cells) anchoring the frontier. But a clear-eyed read of the field warns the wins are fragile: scored honestly, deep models often don’t clearly beat trivial ones — “naive baselines that predict the dataset mean are stubbornly hard to beat on the wrong metric” — and the real test, generalizing to unseen perturbations and unseen cell types, “is brutal.” Tellingly, the Virtual Cell Challenge drew 5,000+ teams from 114 countries because the community hasn’t agreed how to evaluate this yet. Why it matters for the lab: our virtual-cell ambitions (ProtiCelli) live or die on the hard split, not the demo — a discipline worth importing wholesale.

🔬 Image-based profiling reaches for assay-agnostic foundation models

Cell Painting — cheap, single-cell-resolution morphological profiling — is having its foundation-model moment. A 2026 review traces the shift from CNNs to self-supervised vision transformers and transformer-based segmentation (CellSAM, CellViT), while a confounder-aware model trained on 13M+ images across 107k compounds reports state-of-the-art mechanism-of-action and target prediction even on unseen compounds. The recurring theme — echoed in a Cell Systems piece on “compositional” foundation models — is accessibility and generalization: most models assume the 5-channel Cell Painting panel and stumble on other microscopy modalities, so the field is pushing toward assay-agnostic, open, benchmarked models. Why it matters for the lab: that’s exactly the open, FAIR, model-sharing world our BioImage Model Zoo and BioEngine were built for.

The through-line: capital and scale are arriving fast, but the honest tests — baselines, unseen splits, cross-modality transfer — are where it’s actually decided. Prediction is cheap; generalization is the moat.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.