Claude Runs BigHat's Lab Through MCP: 1,232 Antibodies Tested in 10 Weeks for Drug Optimization
Synopsis
Turning an antibody into a drug means tuning binding, stability, aggregation and other properties that pull against one another, and lead optimization often takes months to years. Anthropic and BigHat had one researcher and Claude operate BigHat's automated lab via MCP to optimize one VHH and one single-chain Fv antibody across seven design rounds. In 10 weeks they synthesized and comprehensively tested 1,232 antibodies; the most improved variants of both bound their targets up to nine times more tightly, one reaching sub-nanomolar affinity, stayed stable in solution at temperatures up to 18°C higher, and aggregated far more slowly.
Interpretation
Antibody lead optimization moved from manual round-by-round trial and error to a closed design loop run jointly by an AI agent and a scientist on an automated lab. Such programs often spend months to years in lead optimization, and human engineers usually change only one or two amino acids at a time. Across two binder families, seven design rounds produced and comprehensively tested 1,232 antibodies in 10 weeks, each round shaped by the previous cycle's data; BigHat reports about one week per round versus at least 3-5x longer for most similar lab cycles.
The agent did not brute-force search; it read the sequences first and reasoned like an experienced antibody engineer. Claude started with only a single antibody sequence per target, its baseline measurements and the project goals. Claude recognized both starting molecules as members of families described in the patent literature and spotted that reversing a single difference from a published ancestor could significantly enhance binding for one starter, which BigHat confirmed in the first design round.
Heavily mutated designs still bound, indicating the agent's calls rested on seeing every result in context round after round. Human engineers usually change one or two amino acids, whereas about 150 of Claude's designs sat 10 to 13 mutations from the starting molecule. In one round, every one of the 65 heavily changed VHH designs still bound its target, with each call relying on deep reasoning over all prior results rather than specialist models alone.
A lab data layer let the agent trace each design to its sample, instrument and settings, and correct a default instrument software's misjudgment. Lab instruments often run black-box software with proprietary inputs and produce disjointed output files. The default software of a gold-standard thermostability instrument made incorrect calls on antibody unfolding and aggregation, which Reccy flagged as misleading; Claude proposed changing the heating protocol by varying protein concentration and heating rate, and BigHat's results supported that model.
Perspective
The demonstrated scope is one VHH and one single-chain Fv antibody, two binder families, seven design rounds and 1,232 antibody tests within 10 weeks, all inside BigHat's automated lab and Reccy data control layer. The collaborators describe this as a pilot for a broader therapeutics initiative aimed at harder problems with less decided up front.
The claim of fastest time to completion among similar experiments is the collaborators' own account rather than an independent comparison; the nine-fold binding gain and up-to-18°C stability gain belong to the most improved variants of the two antibodies, not to all designs. The collaborators also note that the bottleneck for frontier models is shifting to data, the lab and the harness.