The Pulse
Google Agents Reach 42.1% Hit Rate on Alpha-Thrombin Aptamers
Google Research says AI agents raised confirmed aptamer hit rates against alpha-thrombin from zero to 42.1% across three experimental rounds. The company also introduced VERID, a digital twin for checking and tracking robotic lab protocols,
AI.info Team · Written with AI ·
Google’s AI-guided aptamer experiments moved from a 0% hit rate to 42.1% against alpha-thrombin in three design-and-test rounds, the company said October 16. Aptamers are short DNA molecules selected for their ability to bind a target; the result is a laboratory binding measure, not evidence of a treatment or clinical benefit. Google Research presented the experiments alongside VERID, software that gives AI agents a live model of a robotic workcell before they issue lab commands.
Alpha-thrombin rose from zero to 42.1%
The first alpha-thrombin round produced no confirmed binders. Google’s researchers traced the failure to fixed primer overhangs that disrupted folding in 48% of the sequences. Agents then introduced clamped flanking stems in the second round and diversified the candidates in the third, lifting confirmed on-target hit rates to 26.0% and then 42.1%, with 98.5% specificity.
The myoglobin campaign shows why the measurement matters. When agents favored raw binding affinity in round two, 63% of their designs contained G-quadruplex motifs. Those molecules often stuck to off-target proteins or the magnetic beads used in the assay: the true, bead-corrected hit rate fell from 10.2% to 3.7%, even as raw pull-down increased. Google says the agents adjusted their approach after the experimental readout exposed the mismatch.
“Agents need real experimental readouts to refine their hypotheses as computational models do not capture the full picture of what happens in a wet-lab reaction and cannot replace empirical validation,” write Pooja Rao, Petar Sirkovic and Annalisa Pawlosky of Google. The aptamer cycles took seven to 10 days each, using pooled DNA synthesis, selection against target proteins on magnetic beads and sequencing to return data for the next round.
VERID tracks what the robot can actually do
Google’s VERID framework—short for Verified Infrastructure for Discovery Laboratories—addresses a separate problem: an AI-generated protocol can be valid code yet impossible or unsafe to carry out. Its live digital twin tracks the workcell in three dimensions, including robot geometry, well volumes, reagent identities and available pipette tips. Agents can query that state, simulate a protocol before execution and receive a block on commands that do not fit the equipment or remaining supplies.
The system also records synchronized video, instrument activity and execution logs. In tests using dye and water rather than biological samples, the records helped researchers diagnose uneven mixing caused by dispensing too close to the bottom of a well. In another test, the digital twin let an agent resume after a plate-reader failure without repeating completed pipetting. Google reports that a separate dye-reconstruction campaign improved accuracy more than fivefold over repeated runs.
Protein design shows a second kind of result
A third project applied agent-guided design to fibroblast activation protein alpha, a receptor overexpressed in solid tumors. Google says the top designs produced binders with single-digit nanomolar affinity after a few experimental runs. In the computational workflow, removing the local binding pocket’s surrounding structure during folding made each cycle more than eight times faster without reducing structural accuracy, according to the company.
In an equal-compute comparison, a human–agent team nearly doubled the in-silico hit rate of a scientist working alone. The scientist set the high-level strategy while an agent handled tactical execution. Google describes that result as computational; it is distinct from the wet-lab binding result, and the company says its next step is to connect protein discovery with VERID’s traceable automation.
Human approval remains part of the process
Google’s work does not describe a lab running without human control. The company says people review and authorize designed sequences, verify protocols and supervise robot runs; it also says every sequence receives biosecurity screening and every protocol passes a simulated dry run before reagents are used. Its test systems were deliberately selected to be controlled and non-infectious, without gain-of-function risk.
Google identifies unresolved tasks: ensuring agents optimize for true target binding rather than experimental shortcuts, checking multi-step protocols and monitoring agent reasoning as well as execution traces. The reported campaigns show that agents can learn from physical measurements, but the integrated system for protein discovery and robot execution remains a stated next step, not a completed autonomous laboratory.