Chan Zuckerberg Initiative builds an AI that reasons like a cell

Scientists trying to use AI for serious biology have always run into a brick wall. Standard chat models can summarize PubMed papers all day, but hand them raw genomics data and open-ended causal questions, and they hallucinate. Meanwhile, actual wet-lab experiments cost a fortune and take months.

The Chan Zuckerberg Initiative (CZI) unveiled rBio, the first reasoning model trained directly on virtual cell simulations rather than just published text or raw lab sheets.

Teaching an LLM to think like a cell

CZI has spent years building virtual cell platforms like TranscriptFormer, which models cell types, development stages, and complex gene regulatory pathways. The problem? These systems output dense, raw numerical data that require dedicated bioinformatics teams just to parse.

The rBio team, led by CZI Senior Director of AI Theofanis Karaletsos and Senior Research Scientist Ana-Maria Istrate, took a clever shortcut. They took an off-the-shelf base LLM and effectively used TranscriptFormer as its private tutor.

Instead of memorizing papers, rBio learned from simulated cellular perturbations and co-expression patterns. Researchers can now literally ask it in plain English: "If I suppress gene A, will gene B activity ramp up? Will the cell tip from healthy to diseased?"

Ditching binary rewards for biology's gray areas

Training reasoning models like OpenAI's o-series usually relies on clear right-or-wrong feedback, like a math puzzle or Python script. But cell biology doesn't deal in clean binaries; it lives in conditional probabilities and fuzzy outcomes.

To fix this, the engineers revamped reinforcement learning. Instead of dishing out binary rewards, they scaled feedback based on how close the model's hypothesis leaned toward biological reality.

The payoff was real: on the PerturbQA benchmark, rBio beat out models like SUMMER (presented at ICLR 2025) and generic baselines like Qwen 2.5. Remarkably, by merely grasping gene co-expression patterns, rBio managed zero-shot reasoning on gene activation and silencing tasks it was never explicitly trained on.

The codebase is now up on GitHub, with a companion bioRxiv preprint and a quick start on CZI's virtual cell platform.

Why it matters

For lab researchers and biotechs, rBio acts as an interactive, zero-cost pre-flight check. Before blowing thousands of dollars on physical reagents to test a hunch about Alzheimer’s pathways, scientists can pressure-test gene knockdowns inside a conversational interface.

We think CZI's real breakthrough isn't just another biology model; it's the architectural blueprint. High-quality wet-lab data is notoriously siloed and expensive. By using specialized simulation engines to train conversational reasoning models, CZI bypassed the biological data bottleneck entirely.

The catch? For now, rBio's domain is mostly confined to gene perturbation. Hallucination guardrails also need to be ironclad when dealing with molecular biology, because a convincing yet flawed hypothesis can still burn massive lab budgets. But turning cryptic digital simulations into a conversational peer is clearly the interface biology has been waiting for.