The pace at which biomedical discoveries translate into actionable health insights has long been throttled by fragmented, repetitive laboratory workflows — a bottleneck that may now have a credible technological answer. A general-purpose AI agent capable of independently executing a wide spectrum of research tasks, without domain-specific fine-tuning, represents a meaningful structural shift in how biomedical science could be conducted at scale.
The system, called Biomni, was developed and benchmarked across 25 biomedical domains. Rather than relying on predefined task templates, Biomni deploys a large language model architecture paired with retrieval-augmented planning and code-based execution. A dedicated action-discovery module mines tools, protocols, and databases from thousands of published papers to construct a dynamic, unified agentic environment. Benchmarking demonstrated strong generalization across heterogeneous tasks including causal gene prioritization, drug repurposing candidate identification, rare-disease differential diagnosis, microbiome composition analysis, and molecular cloning design — all without task-specific tuning. Real-world case studies extended these capabilities to protein stability optimization and direct orchestration of physical wet-lab instruments.
What distinguishes Biomni from narrower biomedical AI tools is its architectural ambition: most existing systems are trained for singular tasks — predicting protein structures, identifying drug-target interactions, or parsing clinical notes. A system that composes multi-step research workflows dynamically, spanning computational and physical laboratory environments, has not previously been demonstrated at this breadth. Published in Science, this work carries institutional weight, though several critical caveats apply. The benchmarks are largely self-constructed or drawn from existing datasets, and independent external validation of real-world discovery quality is absent. Accuracy metrics for high-stakes outputs like rare-disease diagnosis remain to be tested in clinical settings. Nonetheless, if the generalization claims survive independent replication, Biomni-class systems could meaningfully compress early-phase discovery timelines — particularly for rare diseases where human research bandwidth is chronically insufficient.