AI agents move onto the wet-lab bench: three closed-loop rounds lift α-thrombin hit rate from 0% to 42.1%
Synopsis
Google Research has connected AI agents to wet-lab liquid-handling robots so that design, experiment and sequencing readout form a closed loop. Previously agents could only do computational design while real experimental readouts lagged. Using DNA aptamers as a testbed, agents revised their own rules after two failed rounds, raising confirmed on-target hit rates to 42.1% for α-thrombin and 33.5% for myoglobin.
Interpretation
AI agents no longer stop at computational design; through liquid-handling robots they execute wet-lab experiments and feed sequencing readouts into the next design round, closing a design-test-analyze loop. The prior condition was that agentic hypothesis generation and in-silico design advanced rapidly while real wet-lab experiments did not keep pace, making experimental readouts the tighter bottleneck. Three rounds each on two protein targets of differing difficulty, myoglobin and α-thrombin, reached confirmed on-target hit rates of 33.5% and 42.1%.
Inside the loop the agents read the previous round's data, identify a failure mode and rewrite their own design rules rather than following a fixed pipeline. α-thrombin's first round yielded 0% binders because fixed primer overhangs disrupted folding in 48% of sequences. Clamped flanking stems in round two raised the hit rate to 26.0%, and diversification in round three reached 42.1% at 98.5% specificity.
For protein binder design, a scientist sets high-level strategy while an agent drives tactical execution. De novo design requires generating backbones, designing sequences and re-folding complexes to predict binding, and re-folding consumes most GPU compute, scaling quadratically with target length. Excising only the local binding pocket during folding accelerated each cycle more than eightfold without losing structural accuracy; at equal compute, human-agent collaboration nearly doubled the in-silico hit rate of a scientist working alone and found hits across a broader range of target sites.
Perspective
The report covers three efforts released by Google Research: VERID, a semantic digital twin of the laboratory workcell; closed-loop agentic molecular discovery prototyped with aptamers; and agent-guided epitope excision for de novo binder design against FAP. The aptamer loop completed three design-test-iteration cycles on each of two targets, myoglobin and α-thrombin; protein design produced single-digit nanomolar binders in a few experimental runs. The stated next step is to combine VERID's traceable automation with closed-loop learning for protein discovery.
α-thrombin and myoglobin each required overcoming a different wet-lab failure mode, so loop performance depends on the target. In myoglobin's second round, prioritizing raw binding affinity enriched G-quadruplex motifs, dropping target specificity from 75.0% to 19.1% and the bead-corrected true hit rate from 10.2% to 3.7%, recovering only after rule revisions in round three. Protein loops add biological expression, creating bottlenecks on both computational and wet-lab sides; ensuring agents optimize true target binding rather than experimental shortcuts, verifying multi-step protocols, and continuously monitoring agent reasoning and execution traces remain open problems.