Google Research's lab agents chased sticky DNA binders until specificity fell to 19%

Google Research says AI agents running its robotic wet lab fell into a classic trap in an October 16 write-up: chasing raw binding strength, they filled a round of DNA aptamer designs with "sticky" sequences that grabbed almost anything, and specificity for the target protein dropped from 75.0% to 19.1%. In the next round the agents changed their own design rules and recovered.

The episode is the most concrete part of a broader report on letting software agents plan, run and learn from real lab experiments rather than only design molecules on a computer.

How the sticky round happened

The test bed was aptamers, short DNA strands that fold into shapes able to bind a protein. They are cheap to make and quick to read out by sequencing, so a full design, test and analysis cycle takes 7 to 10 days, according to Google. The agents ran three such cycles against two targets, the muscle protein myoglobin and the blood clotting enzyme alpha-thrombin.

On myoglobin, the agents' second round leaned heavily on G-quadruplex motifs, a guanine-rich fold that made up 63% of the designs. Those strands bound strongly, but much of that binding was to the wrong things, including off-target proteins and the bare beads used in the assay. Once bead binding was subtracted, the true hit rate fell from 10.2% to 3.7%.

For round three, the agents capped promiscuous motifs, subtracted bead binders from their scoring and moved to scaffolds that were not G-quadruplexes. Google reports that the on-target hit rate rose ninefold to 33.5% and specificity came back to 72.4%.

A different failure on the second target

Alpha-thrombin went wrong for another reason. In round one, fixed primer overhangs, the sequence tails needed for the lab chemistry, disrupted folding in 48% of the designs, and the round produced no binders at all. Clamping those tails into stems fixed the problem: hit rates reached 26.0% in round two and 42.1% in round three, at 98.5% specificity.

Google frames both cases as the agents learning from their own mistakes, but they also show the open problem it names: making sure an optimizer rewards real target binding rather than a shortcut in the experiment.

Keeping a robot from doing something impossible

The lab side runs on a system Google calls VERID, a digital twin of the robotic workcell built on the open-source PyLabRobot framework. It tracks arm positions, well volumes, reagents and pipette tips, simulates a protocol in seconds and blocks commands the hardware cannot carry out.

In one case it refused an eight-channel move that was too wide for a narrow trough. Forced to rethink, the agent designed a diagnostic run that exposed a small bias from liquid clinging to tip walls. In a self-play task where it reconstructed hidden dye mixtures from camera images, the agent improved its accuracy more than fivefold, Google says.

A separate protein design project targeting FAP, a receptor common in solid tumors, found that a scientist directing an agent nearly doubled the computational hit rate of a scientist working alone at equal compute. Google did not publish the exact binding values or binder counts, and it says every designed sequence went through biosecurity screening and human review.