rationale:

The modal path (37%) is an AI system designing experiments or molecules that humans then execute. Current frontier biomedical AI heavily follows this "AI proposes, humans run" paradigm. For example, Robin generated therapeutic candidates for dry AMD and defined follow-up experiments that humans ran A multi-agent system for automating scientific discovery, A multi-agent system for automating scientific discovery, and Co-Scientist produced target hypotheses validated by human collaborators Accelerating scientific discovery with Co-Scientist, A multi-agent system for automating scientific discovery. Even the most advanced recent result—Anthropic's unrandomized mouse-efficacy study for muscular dystrophy—relied on Claude coordinating 11 contract labs to do the physical wet-lab work Early progress on accelerating drug development with ..., https://www-cdn.anthropic.com/52e1f5d50c47ca3818f56704c27cf19e59b62e77.pdf. Because a Nobel-caliber result requires either a fundamental discovery or first clear evidence of human efficacy, human-executed laboratory and clinical trials remain the unavoidable bottleneck, making this the most natural classification for the decisive work. These structural validation requirements favor workflows that incorporate human execution rather than attempt to eliminate it.

Dataset mining is a strong secondary pathway (26%). A major biological mechanism could be discovered from existing data long before a newly designed treatment demonstrates human efficacy. Leveraging existing biobanks, health records, or genomic archives provides a much faster route to a prize-worthy breakthrough. This makes mining existing evidence particularly attractive when prospective laboratory and clinical validation are the main constraints. Kosmos has already demonstrated autonomous discovery from datasets, reporting novel contributions and reproducing unpublished findings Kosmos: An AI Scientist for Autonomous Discovery. Similarly, Claude’s ART genome-mining campaign found a new CRISPR-like enzyme family, showing how agents can notice unexpected biological patterns Autonomous AI agents discover reverse transcriptases with .... If an AI system surfaces a profound disease mechanism or repurposing signal from existing archives, the decisive work will be classified here, even if humans subsequently run clinical trials to confirm it.

General systems are also increasingly adept at autonomously directing purpose-built specialist AI tools, keeping specialist orchestration highly relevant (20%). Protein and antibody design naturally combines general reasoning agents with dedicated generative and structural systems. For example, a recent Claude campaign autonomously operated structural-design tools to produce confirmed binders https://www.anthropic.com/research/Claude-accelerates-protein-design, and it recently orchestrated 12 tools to create scorpion-toxin binders https://www.anthropic.com/research/claude-scorpion-toxin-binder-design. A first-in-class biologic designed through this workflow could resolve here, provided the general system is credited with the central idea and the specialist tools act strictly as subordinates. Existing specialist-only drug successes, like rentosertib reaching Phase 3 Insilico Initiates Phase III Clinical Trial for Rentosertib, Its AI ..., A generative AI-discovered TNIK inhibitor for idiopathic ..., do not satisfy the criteria for a general-system-produced result.

The remaining probabilities are split between pure literature reasoning (9%) and closed-loop autonomous labs (8%). Pure reasoning over literature and theory is a dominant path in mathematics, but medicine strictly demands empirical confirmation; purely theoretical synthesis is unlikely to be recognized as the decisive work without accompanying biological evidence. This validation requirement limits the probability that literature reasoning alone supplies the decisive contribution. Meanwhile, closed-loop autonomous labs remain constrained. While GPT-5 successfully drove Ginkgo's automated lab to achieve a 40% reduction in protein-production costs and a 27% higher titer Using a GPT-5-driven autonomous lab to optimize the cost ..., GPT-5 lowers the cost of cell-free protein synthesis, this demonstrates process optimization rather than transformative biological discovery. Crucially, human clinical efficacy cannot be achieved entirely within a closed-loop lab, relegating this pathway to preclinical discoveries that happen to be hailed as prize-level on their own.