AI for Science: Drug Discovery, Protein Design & Autonomous Research

AI for Science: from research tool to research loop
April's two biggest infrastructure bets — Biohub's $500M virtual cell push and DOE's Genesis Mission lab at PNNL — make AI the experimenter, not the tool inside the experiment.

For two decades AI in biology meant a sharper instrument inside a human-led research loop: a sequence aligner, a structure predictor, a search ranker. April 2026 closed that loop. On April 29 Biohub announced a $500M, five-year Virtual Biology Initiative, anchored by NVIDIA and the Allen, Arc, Broad and Sanger institutes — explicitly aimed at a predictive model of the human cell where mutations and drugs are tested in silico before anything goes near a wet lab. A day later DOE commissioned AMP2 at Pacific Northwest National Lab and gave Ginkgo Bioworks $47M for the larger M2PC successor — 97 robots and 100+ instruments by 2030, running anaerobic microbial experiments around the clock with humans largely out of the building.

The two bets are complementary. Biohub goes after the data foundation: cellular biology has roughly billion-cell datasets, predictive cell models likely need orders of magnitude more, and nobody yet knows the scaling-law slope for biology the way OpenAI knows it for text. PNNL and Ginkgo go after throughput: if you need much more data, you need labs that generate it without humans handling pipettes. Robots run the experiments models propose; models train on the data robots produce.

Pharma has been quietly paying for the same thesis. Eli Lilly's March 2026 deal with Insilico Medicine — up to $2.75B, $115M upfront — bought not a single molecule but the pipeline: 42 AI models that pick targets, design candidates, and predict trial success. Insilico's lead asset reached clinical trials in 18 months against the standard four to six years. It's the first nine-figure-upfront pharma deal where the asset is a discovery process, not a compound.

The "autonomous scientist" is real now, if narrow. AI-Scientist-v2 runs the full cycle — literature review against Semantic Scholar, novelty filter, agentic experimental search, paper drafting with figures and citations — in roughly 15 hours and $140 per cycle, on a paper that passed peer review. A Russian consortium claims a paper accepted at an A-tier IT venue with 90% of the work automated. OpenAI has pre-announced an "autonomous AI researcher" for fall 2026, pitched as a way to bootstrap the next model generation. The risk Gonzo-ML flagged in March still applies: LLM-written literature summaries get scraped back into training data — citogenesis — and look authoritative without any human ever checking.

One demonstration is harder to file. An Australian ML engineer with no medical training designed and produced a personalized mRNA cancer vaccine for his dog using ChatGPT and AlphaFold: tumor sequencing, prediction of which tumor-specific peptides the immune system would see, and a manufactured dose. The dog responded therapeutically. The personalized cancer-vaccine pipeline that took multi-million-dollar labs a decade is now reproducible by one motivated person with off-the-shelf AI tools, and there is no safety frame for that yet.

Stanford's AI Index 2026 puts numbers on what shifted in twelve months. AI agents complete 66% of OSWorld tasks — a generic computer-use benchmark — against 72% for humans, up from 12% a year ago. SWE-bench Verified, the coding benchmark stuck near 60% in early 2025, is essentially saturated. Claude Opus 4.6 and Gemini 3.1 Pro both cross 50% on Humanity's Last Exam, where o1 scored 8.8% eighteen months earlier. The Foundation Model Transparency Index dropped from 58 to 40 over the same window: the labs got more capable and less open at exactly the same rate.