Toward chemical superintelligence
AI is reaching superhuman ability in domains like coding and math, where reinforcement learning with verifiable rewards (RLVR) can capitalize on the near-infinite scalability of compute and reliable verification. Superhuman chemical intelligence would transform drug discovery, agrochemicals, materials science, and the broader specialty chemistry industry, but these domains do not obviously offer such a reward signal.In drug discovery (our focus), it takes a decade to learn whether a single drug is safe and effective. Yet drug discovery is not one action but rather thousands of smaller actions (choosing which molecule to test, predicting properties, setting up a reaction, etc.) that organize hierarchically and have verification latencies that vary by orders of magnitude.Viewed this way, a path to chemical superintelligence emerges: generate data at all levels of the hierarchy, and use progress on the short-horizon tasks to guide improvement on the long horizon.
[data]
[ai]
[software]
[robotics]
[human in the loop]
[impact]
A dataset that scales with the industry
Frontier models for chemical prediction and reasoning
The operating system for chemical discovery
Building for the physical AI era
Learning from and improving the world’s best chemists
AI in service of human health
We generate vast datasets through the process of chemical optimization in drug discovery, including data on physical properties, synthetic chemistry, and biologic function. Our data consortium pools precompetitive data across partners, and our lab captures the inputs and outputs behind every decision, measurement, and action in the reactions we run.
Our Beacon models have consistently taken first place in industry-leading benchmarks. We combine chemistry foundation models with physics-based simulation and physiological modeling to predict properties from the molecular scale through human dose. Our agents use Beacon to make decisions and learn in RL environments built from real drug programs.
Software underlies everything we do. Distributed systems help us scale model training and low-latency inference. Real-time scheduling and logistics ensure that each synthesis step is done at the right time by the right person or machine. Tooling for scientific data capture and real-time chemical design environments help us run efficiently.
Automation in chemistry requires more than high-throughput pipetting. Our labs are instrumented to generate and learn from the data needed to automate the long tail of drug discovery chemistry. Our approach accelerates every synthesis and scales naturally as AI moves to the physical world.
Experienced wet-lab scientists guide our experimental work and data collection. Our human-in-the-loop AI systems combine model predictions with expert judgment to help chemists make better decisions. Over time, the expertise developed in our labs becomes encoded in the models, tools, and automation that support them.
Our technology is deployed across dozens of active drug programs spanning oncology, neuroscience, immunology, inflammation, cardiometabolic disease, rare disease, and other areas of high unmet need. Every prediction and experiment helps our partners make better decisions today while teaching our models how to discover better medicines tomorrow.
