Medra's AI scientist solved a critical antibody binding problem (zero to 70%+ success) by proposing a simple vortexing step—something human researchers dismissed as too basic to test. The company opened a 38,000 square foot autonomous lab in San Francisco with ~100 robotic arms…
The antibody binding result was zero. Not low — zero. The antibodies simply were not sticking to the target protein, and nobody knew why. Michelle Lee's AI scientist at Medra considered the usual explanations, ran the usual tests, and then proposed something unusual: add a vortexing step in the middle of the protocol. Binding jumped from zero to more than 70 percent. Core Memory
The discovery was not a new molecule. It was not a novel mechanism or a surprising insight about protein structure. It was a lab mixer — used at the right moment — a technique so mundane that most scientists would have dismissed it as obvious in retrospect. The AI had tried what humans considered too basic to test.
Medra opened its 38,000-square-foot autonomous lab in San Francisco this week. About a hundred robotic arms run continuously across three floors of weight-bearing concrete, each arm positioned beside a different scientific instrument — centrifuges, incubators, pipetting stations — all of them operated by software that reads like a trained human. Cameras mounted near each gripper log the exact angle of every pipette tip, the exact depth of insertion, the timing between reagent additions. With humans, that information is tacit: an experienced scientist builds intuition over years, and when they leave, their knowledge leaves with them. Medra's sensors are among the first systems to put it on the record. Core Memory
"We adapt general robots for the reality we live in," Lee says. The arms are sourced from the same manufacturer that supplies Toyota factories.
The second layer is the AI scientist: a software agent that reads experimental results, identifies what is going wrong, proposes protocol changes, and rewrites the protocol itself — with no automation engineer in the loop. A customer brought a binding problem. The AI narrowed it to two hypotheses, designed a test to distinguish them, and proposed the vortexing modification. The result was not incremental. It was the difference between a failed experiment and a working one.
Only about 5 percent of the instruments on a scientist's bench currently fall into the "can be automated" category, according to industry estimates. The rest — centrifuges that need balancing, pipettes that need gripping at a specific angle, protocols that depend on timing a human hand can feel but a machine cannot — were designed for hands. Medra's claim is that its software can extend automation to the other 75 percent, using computer vision and manipulation models to adapt to instruments labs already own. Core Memory
Whether it can deliver on that claim is the open question. Lab automation and AI scientists have been overpromised for two decades. But the physical infrastructure is real: 38,000 square feet, 100 arms, five customers with experiments scheduled, and a Genentech partnership for early drug discovery announced last month. The company raised a $52 million Series A in December, bringing total funding to $63 million, led by Human Capital with participation from Lux Capital, Neo, and NFDG.
Lee grew up in Taiwan, came to America at 14, and spent 2015 interning at SpaceX. She was on track for a faculty position at NYU when AlphaFold 2 was released in 2021, and she began thinking about why protein folding had become solvable. The answer, she concluded, was fifty years of structural data. Drug target validation and antibody design lack that data — and the only way to generate it is to run more experiments than any team of humans can manage. In 2024, she rebuilt the hardware and software from scratch, designed to be reconfigured for each customer rather than sold as a fixed product. Core Memory
The first Medra customer signed a six-figure contract before the company had a robot in its lab — on the strength of a PowerPoint and borrowed photographs of someone else's robotic arm. The team was Lee and one other person. Core Memory
More than 85 percent of Medra's customers arrive with requests the company has never fulfilled before. Because the software and hardware layer is consistent across protocols, reconfiguring from one setup to a hundred does not require rebuilding from scratch. Medra went from no arms in the building to a hundred arms running antibody binding in under three months. Core Memory
The TSMC comparison is the one Lee uses unprompted. TSMC manufactures the chips that make it possible for chip designers to exist. Medra wants to be what makes it possible for a drug discovery company to run experiments without building its own lab. The question is whether that manufacturing analogy holds — whether the bottleneck in drug discovery is really throughput, and whether automating the physical lab is the right place to attack it.
The arms were still running when a reporter visited. They will still be running tonight, and tomorrow, on a schedule that does not stop at five and does not take weekends. The jobs queue and clear. The protocols complete. The small courier robot continues its circuit — tip rack to plate to centrifuge and back — moving through the room with the same unconcern it showed for the visitor.
"If we could cure cancer, Alzheimer's, infectious disease — we have the ability to do that," Lee says. "We just don't have the throughput."
The open question — the one that will determine whether this is a real inflection point or another chapter in the long history of overpromised lab automation — is whether the AI scientist's judgment is sound. The vortexing result is a single data point. Medra has not published the experimental logs. It has not made the decision trail available for inspection. The company says the logs exist and the process worked. The outside world has only their word.
That is the go-look test. An AI that can find answers humans missed by trying what humans considered too basic to test — that is worth looking at closely.
Story entered the newsroom
Assigned to reporter
Research completed — 6 sources registered. Medra has 100 robot arms running 24/7 in a new 38,000 sq ft San Francisco facility. An AI scientist identified that adding a vortexing step could jump
Draft (968 words)
Published (980 words)
@Rachel — drop story_11639. Three stories stuffed into one wire roundup: Novo's pediatric oral GLP-1 (skeletal data, filing H2 2026), Altimmune's M offering (textbook dilution move), and the Merck-Google Cloud B AI tie-up (needs pharma-native framing, the Endpoints item is too thin). None of these will land right compressed like this. Reject the bundle; we can revisit each one individually if a standalone source actually delivers the goods.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Biotech & Life Sciences · 1d ago · 5 min read
Biotech & Life Sciences · 1d ago · 2 min read
