Quine: deciding what a world model of biology should learn next

What Quine is

Biology runs on slow loops. A scientist proposes something, waits weeks for an experiment, and spends scarce samples on the answer even when it’s no. Quine is Microsoft Research’s attempt to make those loops shorter. It has two parts that are built together: a world model of biology that learns shared representations across genomics, proteins, chemistry, RNA and cell state, and imaging, and a harness that connects that model to scientific tools, the literature, the wet lab and the scientists using it.

The public announcement describes the result we care about most. With collaborators at the Broad Institute, Quine ranked thousands of compounds by their predicted ability to shift pancreatic cancer cells between states. In the lab, the top-ranked compounds produced the largest intended shifts, and narrowing the search took one weekend, potentially saving months of experimental work.

What a research product manager does here

The model is a large language model post-trained across five biological modalities. Our training stack is mid-training followed by reinforcement learning with verifiable rewards (RLVR). My job is the set of decisions that sit around the training runs: which capabilities matter, how we’ll know we have them, what data and compute it takes, and when a checkpoint is good enough to put in front of a scientist.

1. Defining capability

I own the capability roadmap. For each model iteration I translate four inputs into capability targets, release gates and training investments:

• Research priorities from the scientists building the model.
• Usage signal from the labs and partners working with the model.
• Competitive analysis. I track Google’s AI co-scientist, AWS’s scientific discovery work and Anthropic’s science efforts, and run external models through our own eval pipeline so comparisons happen on our tasks instead of someone else’s launch post. That analysis set our data acquisition strategy and the synthetic data strategy behind 4+ shipped capabilities.
• Eval results, which decide whether a target was hit or needs another iteration.

Capabilities need data, so I also own acquisition and onboarding for training and evaluation data across all five modalities: intake, licensing and compliance review, and quality gates across 40+ internal and external datasets.

2. Leading evals across mid-training and post-training

Evals are how a research team agrees on what “better” means. I lead them end to end:

• The team’s first RLVR program. I designed the verifiable-reward task set, the evaluation framework and the learning curriculum, and found the training-data and preprocessing gaps holding it back. Fixing them improved held-out task accuracy 12%.
• Partner evaluation programs. I ran evals with research labs and five Fortune 500 pharma R&D partners across drug discovery and preclinical workflows like cell type annotation and perturbation prediction. Real-world failure modes became new mid-training data and RL tasks, and benchmark pass rates rose 15% across three model iterations. I’ve negotiated 10+ evaluation partnerships with academic institutions and Fortune 500 companies so far.
• Promotion gates. I set the model-readiness criteria that turn benchmark and held-out results into go or no-go decisions, agreed across 30+ researchers, plus the data and eval dashboards they use to catch regressions.
• Compute. I direct 50,000+ GPU-hours across mid-training and RL, prioritizing experiments, writing their success criteria, and moving compute toward the capability gaps the evals expose.

3. Getting it to scientists

A model scientists can’t use doesn’t change any science. I’m the product lead for the multi-agent research workbench that sits on top of the model. It orchestrates literature review, data analysis and agentic workflows into wet-lab-ready experimental plans. I ran the workflow research that decided which connectors and skills we built, then onboarded and supported the scientists using it. It has reached 100+ scientists across four labs and cut the time from research question to experimental plan by 30%.

I also lead external deployments. The research system went to four academic labs and the model went to four enterprises, each with telemetry, evals and benchmarks I defined with the researchers involved. Those feedback loops reduced the time to generate and rank candidates by 15%. On the wet-lab side I run the validation loop with four contract research organizations, so model predictions get tested in assays and the results come back as training and evaluation data.

4. Safety

I’m the program and product lead for biosafety. With policy, research, engineering and Microsoft’s internal safety board, I set up release criteria and a review process now used across Microsoft research teams and 20+ models. That review surfaced a case where model outputs could be combined with public data to enable misuse. It changed a release decision and led to new output filtering.

5. The website

I built the Quine website with Codex. It’s where scientists applied to the first Quine Research Fellowship, with applications open from September 29 to November 2, 2026.

What I’ve learned

The eval is the real product spec. Whoever writes it decides what the model gets better at, so I try to write evals from what scientists actually do instead of from what is easy to grade. Compute should follow evidence: the fastest way to waste 1,000 GPU-hours is to keep training on a target your evals stopped believing in. And scientists rarely ask for a smarter model. They ask for a shorter path to an experiment worth running, which is a product problem as much as a research one.