Data Engine For Autonomous Science: Michelle Lee, Medra

Author: Michelle Lee, Founder & CEO, Medra Collection: UNLOCK-2026-Sessions

Summary

Michelle Lee, founder and CEO of Medra, presents the case for "physical AI scientists" as the missing infrastructure for autonomous science. Medra came out of stealth a few months prior to this talk and is building exactly this: a robotic AI system that can execute biological experiments with the same breadth and adaptability as a bench scientist, while continuously learning from the process data generated during execution.

Lee frames the problem through three historical phases of AI for biology. Phase 1 (decades ago): digitization, where automated high-throughput screening, automated PCR, and next-generation sequencing created large data streams, enabling bioinformatics and early ML in the dry lab. Phase 2 (last 10 years): scale, where deep learning (CNNs, transformers) produced an explosion of biology models — protein structure, gene expression, multimodal foundation models. Phase 3 (current/emerging): intelligence, and eventually super-intelligence. Her argument is that reaching Phase 3 is not inevitable — it requires solving the "data engine" problem: a continuous feedback loop where models generate predictions, those predictions are validated by experiments, and the experimental results (along with rich process metadata) flow back to improve the models.

The core diagnosis is that design and validation are decoupled in current biopharma workflows. AI scientists on the dry side output protocol designs; lab automation on the wet side outputs results and metadata. But the context that links the two — the AI scientist's reasoning, the experimental process observations, the tacit knowledge accumulated by a bench scientist during execution — is lost. Lee draws an explicit analogy to Tesla's data engine (shown in a 2018 Tesla AI Day slide): Tesla's self-driving system could identify inaccurate predictions, flag them for review, and use them to fine-tune and relabel models in a continuous improvement loop. Biological science currently lacks this.

Medra's solution is a physical AI scientist with three core properties: modularity (can use any instrument, including those without APIs, via computer vision and physical manipulation rather than being locked into API-accessible SPS-format instruments), continuous learning (logs all process data — pipette angles, tip depths, tactile signals, images, error states — to enable observability and model fine-tuning), and scalability (can run thousands to hundreds of thousands of experiments in parallel). Medra Lab 1, announced at UNLOCK 2026 and opening that week, is a 38,000 square foot facility in San Francisco with 100+ robots running 24/7. It is already running in production with biopharma customers including Genentech. Lee claims that in three years of development, Medra has automated close to 65% of laboratory instruments, compared to ~5% for 20 years of traditional automation, and projects 85% within one additional year.