Drug discovery is one of the most important problems in the world and one of the least solved. Biology is complex; the useful data is sparse; the work runs in slow cycles of trial and error. Teams spend years and billions of dollars on development, and most of the time the drug still fails.
Brute force will not fix it: more compute and more data have not made biology predictable. Making today’s process incrementally faster will not either, because that process is capped by what a human team can hold in their heads. What works is a different way of doing the science itself, an AI framework that reasons through it end to end, drawing on the published literature and the data and closing the loop with the wet lab. It’s built around what today’s agentic models are good at, and guarded hard where they fail.
The proof of concept is ready and has been tested on several use cases. In a single run it processed an entire disease’s single-cell data, ingested hundreds of experimental facts extracted from the literature, built a connected map of the biology and returned drug-target hypotheses. Each hypothesis came back marked supported or refuted, traceable with provenance to the exact evidence behind it. When the data is this thin, that grounding is what separates an answer you can build on from one you take on faith. The work came out of a collaboration with the Allen Institute, and the first results are submitted to a top-tier journal (Cell). Preprint: https://www.biorxiv.org/content/10.64898/2026.07.01.734821v1.
The system already produces strong results for parts of the drug development pipeline, target validation among them, powered by the data modalities we have built in. The plan is to scale that to every modality and reach what the whole field is chasing: reliably predicting what a clinical trial will do before it runs, with a concrete explanation of the biology of why.
We prove things with hard, contamination-free benchmarks rather than decks. We’ve built a freedom-to-operate benchmark for our internal tool and are working towards a clinical-trial benchmark with mechanistic explanation, because a trial is the one place biology gives a straight answer. An engine validated against that benchmark is more than capable of tackling every stage of drug development, and in time, of showing not just what a trial will do but how to change it.
That’s the mission. Not to make today’s process a bit faster, but to change what is possible.