Virtual Lab
Cancer researchers have public datasets and hypotheses, but getting from a question to a tested, cited answer takes days: someone has to read the literature, design the analysis, poke holes in it, then write and run the
Core trick
Gemini writes survival-analysis code that runs on METABRIC patient data in a managed sandbox; a blind judge scores it against a no-tools baseline.
Description
Cancer researchers have public datasets and hypotheses, but getting from a question to a tested, cited answer takes days: someone has to read the literature, design the analysis, poke holes in it, then write and run the code. An AI chatbot answers in seconds, but it never touches the data, so its numbers are educated guesses.
Virtual Lab works like a real lab, built from 28 Gemini agents with one job each. They study the topic and write cited notes. Three planning meetings run in parallel: in each, a lead scientist asks an immunologist, a statistician and a data expert for ideas, and a critic attacks the plan until it holds up. A lead scientist merges the plans and code checks that the data has every variable needed. Then a person approves the budget, which becomes a hard spending cap. Gemini 4 Argon writes the analysis code, and a Gemini managed agent runs it on 1,980 real breast cancer patients (METABRIC) in a Google-hosted sandbox. The lab writes an answer that separates what its data shows from what published research says, with references. Finally a blind judge compares it with Gemini 4 Argon answering alone, twice, with the order swapped; the lab must win both.
On our demo question (is more cytotoxic immune activity in triple-negative breast cancer linked to longer survival, after accounting for age, stage and grade?), the lab won 8 of 9 blind tests. It found about a 20% lower risk of death per standard deviation of the immune score (hazard ratio 0.76–0.80 in every run), consistent with the literature. A run takes about 2.5 minutes and costs about $0.35.
It is a research tool, not medical advice: questions about an individual patient are turned into research questions about patients in general, and it uses public data only.
README excerpt
# Virtual Lab A lab of 28 Gemini agents that answers a cancer research question the way a real lab would. It studies the topic, plans in three parallel meetings, critiques the plan, asks a human to approve the budget, runs the analysis on real patient data in a Gemini managed-agent sandbox, and then has to beat **Gemini 4 Argon answering alone** in a blind test. One run takes about 2.5 minutes and costs about $0.35. Built at the **Gemini 4 Argon SF Hackathon** (Cerebral Valley × Google DeepMind, 2026-10-25). ## Results Demo question: *"In triple-negative breast cancer, is higher cytotoxic immune infiltration associated with longer overall survival, independent of age, stage and grade?"* Data: 1,980 METABRIC patients, of whom the lab itself selects the 320 triple-negative cases (168 deaths). | | | |---|---| | Blind test (judged twice, order swapped, lab must win **both**) | **won 10 of 11 runs**; 1 split | | Lab's finding, every run | HR **0.76–0.80** per SD of cytolytic score, 95% CI excluding 1, p = 0.002–0.012 | | Question to verdict, live with a human approval click | **150–198 s** (6 of 7 fast-mode runs under 3 minutes) | | Cost per run | about $0.35 (≈120k tokens) | The judges' reason was the same each time: the lab ran the survival analysis on the data, while Argon alone could only reason from the literature and estimate. Different runs planned the analysis differently (complete-case adjustment for stage, n = 227, or stage proxies, n = 293–320), and the hazard rat