Foundation models come for the cell

Agentic Discovery  ·  Nº 51  ·  07 Jul 2026

Editor's Note

Two virtual-cell bets land on the same Monday morning — one open, one Xaira's — and the reference stack for biology AI just got crowded.

Nº 01 · The Lede  —  X  —  Field report

OpenDDE opens all-atom co-folding

Fig. I  X · Filed 07 Jul 2026.

OpenDDE launches as an open-source all-atom foundation model for biomolecular co-folding — the first drug-discovery-engine-scale release with weights, code, and training recipe in the open. Announced by William Hua at ICML, the model targets the same protein-ligand co-folding surface that Isomorphic and Chai keep behind APIs. It's positioned as step one of a stated "open drug discovery engine" roadmap, resetting what the open-weights tier of structural biology looks like.

Read the source →

Nº 02  —  X  —  Cell biology · Funding

Fig. II  X · Filed 07 Jul 2026.

Xaira unveils X-Cell virtual cell

Xaira Therapeutics debuted X-Cell, its first virtual-cell foundation model, positioned explicitly not as a perturbation predictor but as a generalist meant to extrapolate to unseen biology. Bo Wang, promoted to Chief AI Scientist to lead the effort, is anchoring Xaira's platform on cell-scale generalization rather than task-specific benchmarks. Raises the ceiling on what "virtual cell" means in industry — the goalpost moves from prediction accuracy on known perturbations to out-of-distribution biology.

Read more →

Nº 03  —  arXiv  —  Field report

Fig. III  arXiv · Filed 07 Jul 2026.

DELPHAI predicts perturbation heterogeneity

DELPHAI models heterogeneous single-cell perturbation responses by pairing learned cell-fitness scores with gene-space retrieval, moving past the population-average predictions that most Perturb-seq models still produce. Where X-Cell above bets on foundation-scale generalization, DELPHAI's retrieval approach anchors a lighter alternative — narrows the compute gap for cell-response modeling without abandoning heterogeneity.

Read more →

Also Filed  ·  Three Briefs from the queue

Nº 04  —  arXiv  —  Benchmarks · Evaluation

Enzyme specificity benchmarks rebuilt

Enzyme-specificity benchmarks get a hard rewrite in a new arXiv paper arguing current tests leak substrates across splits and inflate accuracy. Establishes cleaner evaluation as table stakes for enzyme-function models, where inflated leaderboard numbers have shaped the last two years of claims — a pattern real-science tests keep exposing across biology AI.

Read →

Nº 05  —  bioRxiv  —  Agents · Infrastructure

Trustworthy healthcare LLM agents

A trust framework for LLM agents in healthcare lays out attestation, auditability, and clinical-workflow guardrails as prerequisites for deployment. Moves the healthcare-agent conversation from capability demos to the accountability surface regulators actually ask about.

Read →

Nº 06  —  bioRxiv  —  Field report

SpliSync corrects long-read splice sites

SpliSync applies a genomic language model to correct splice-site errors in long RNA-sequencing reads, closing an accuracy gap that has kept long-read isoform calls behind short-read gold standards. Moves genomic LLMs from annotation aid to inline read-correction step.

Read →

· · ·

Reply with your discoveries. A human reads them. Forward freely.