Google's wet-lab agents improved after a failed round; humans still ran the lab

Google's aptamer agent fell from a 10.2% hit rate to 3.7%, then reached 33.5%. It caught its own error, while people ran the lab and the robot software handled only dye and water.

Google Research says its AI agents will eventually run laboratory experiments as well as design them. Its blog says experimental readouts are "more of a bottleneck than ever." The most useful result in its new papers is a round in which the agent got worse.

The hit rate for myoglobin, a test protein, was 10.2% in Round 1. It fell to 3.7% in Round 2 before reaching 33.5% in Round 3. The agent lost ground before recovering.

The useful result is the loop. The agent made an error, found it in its own data, and fixed it within one cycle. The cost is low enough to make repetition practical.

People did the lab work, and the robot software was tested on coloured water. The "autonomous wet lab" therefore remains two halves that have not yet been joined.

The aptamer paper covers aptamers, short DNA strands that bind a target protein. When the agents prioritised raw binding signal, 63% of their Round 2 designs carried G4 motifs. These guanine-rich sequences cling to almost anything, and target specificity fell from 75.0% to 19.1%.

The paper says the analyst agent diagnosed the problem without human input. It found that most cross-reactive binders were G4s. A control with bare magnetic beads showed that they were sticky rather than selective.

Round 3 moved 72.3% of the pool onto other scaffolds, restoring specificity to 72.4%.

The blog's "ninefold" gain is measured from the Round 2 trough, 3.7% to 33.5%. From the campaign's starting point, the gain is 3.3 times. Both figures are bead-corrected hit rates among newly designed candidates; neither measures binding strength.

Thrombin, the second target, ran at 0.0%, 26.0%, and 42.1% across the three rounds. The agents found that fixed primer overhangs, the extra DNA tails used to copy sequences, had disrupted folding in 48% of the first round's sequences. Again, the agents found a flaw in their earlier designs.

The economics explain why Google chose aptamers. Each cycle took about ten days for three bench steps. The cost was roughly $3 per candidate for pooled oligo synthesis, meaning DNA bought as one pool, plus reagents.

About 1,000 candidates ran in each round, and 4,400 sequenced constructs were tested in all. At $3 each, those constructs cost about $13,200 in oligos and reagents. The figure leaves out labour, sequencing runs, and compute. The loop is cheap because the DNA is bought as a pool and read by sequencing.

People did the bench work. The paper states that "between rounds, human researchers executed the wet lab pulldown, PCR amplification, and sequencing workflow." PCR is the step that copies DNA. The humans then returned read-count matrices to the agent pipeline.

They held assay conditions constant across the three rounds. The analysis loop was the agent's own because it chose the Round 3 rules; the bench work was not.

The paper says it cannot separate the gain from the agent itself and the gain from the loop. It lists "non-reasoning control arms" as future work, along with head-to-head tests against evolutionary, Bayesian, and supervised machine-learning baselines.

A cheap loop with a bead control might improve any competent optimiser. The paper does not show that the agent's reasoning was the ingredient that mattered.

The robot half is real engineering. VERID, described in its own paper, is a digital twin of the lab, meaning a software copy. It tracks arm geometry, well volumes, and pipette-tip stock so agents "never send unverified code directly to the liquid handling robots." It caught a pipette that would not fit a trough before any liquid moved.

Its validation used "aqueous dye and deionized water mixtures." The headline result was a more-than-fivefold improvement in reconstructing blinded dye mixtures from camera images. That tests the optics and plumbing.

The binders themselves are modest. Four titrated thrombin constructs built on literature motifs bound at 24.6 to 133 nanomolar, where a lower number means tighter binding. All 25 literature-core designs in Round 3 hit, making up 25 of the 131 hits. The strictly de novo scaffolds that were titrated bound only at the assay ceiling, above 1 micromolar.

Supervised machine learning, the paper notes, reached 1.5 nanomolar by training on nearly a million physical measurements. The agents used about 4,400 and have not approached that affinity. The case for the loop is speed and cost, not binding strength.

Google's epitope excision paper adds a protein result. An epitope is the part of a target that a binder grabs. Cutting each target down to its binding window reduced GPU hours by 88%, sharply lowering the computing bill.

An agent-guided workflow with a human in the loop reached a computational hit rate of 9.33%. The principal investigator working alone reached 5.06%. In yeast display against a protein called FAP, four representative clones bound at 3.0 to 9.1 nanomolar.

Those are computational results and yeast-display results, where yeast cells carry the binder on their surface. The paper lists solubly expressed binders and inhibition assays as work still to do.

The remaining gap is integration, which the blog names as its next step. It wants to combine VERID's traceable automation with closed-loop learning for protein discovery. Proteins must be expressed biologically rather than synthesised, so each protein loop adds cellular or cell-free selection, expression, and folding. The FAP work used three rounds of cell sorting, while the aptamer loop used a sequencing readout.

A selection already read by sequencing can support an agent now, without waiting for robots. The cycle is close to ten days at about $3 per candidate. The aptamer data favour starting there, provided the negative control is built into the score before the agent sees it.

Without bare beads and off-target proteins in the objective, a second round could repeat the myoglobin failure. The same speed will not carry over to protein work, where expression and sorting set the pace.

Google gives no date for joining the two halves. The desk's own call is a first protein campaign through VERID by mid-2027. The blog names it as the next step, and the protein papers already report computational and yeast results. Cycles should run for weeks rather than ten days because expression and sorting are slower than sequencing. The comparison figure is the $3 per candidate recorded here.