Google Research's October 16 post brings together three studies that connect AI agents with physical lab feedback, from aptamer design to robot controls and protein binders. In one campaign, agents revised molecular designs after experiments exposed failures; in a separate protein-design effort, four selected FAP-binding clones showed apparent affinities in the 3.0–9.1 nanomolar range. Those are early research results, not evidence of a treatment.
Experiments changed what the agents designed
In the aptamer study, agents proposed short DNA molecules that bind proteins, while researchers synthesized the designs, screened them against two target proteins and returned sequencing measurements for the next round. The team ran three design-test-improve rounds, each with about 1,000 candidates. A full cycle took roughly a week to ten days.
The share of newly designed candidates that passed the study's bead-corrected on-target binding test rose from 0% to 42.1% for alpha-thrombin and from 10.2% to 33.5% for myoglobin by the third round. The thrombin figure includes descendants of known binding motifs; the rate for strictly de novo candidates was 37.1%. The paper counted an aptamer as a hit only when its target enrichment was at least twofold over the input pool and statistically significant, and when it exceeded a fitted background from bare streptavidin beads. That correction filters out sequences that stick to the assay matrix rather than specifically binding the target.
The path was not a straight climb. On myoglobin, agents initially favored G-quadruplex motifs that stuck to other proteins and assay beads. The measured hit rate fell to 3.7% in the second round as target specificity dropped from 75.0% to 19.1%. After the agents identified the cross-reactivity and shifted designs toward other structures, the third-round hit rate reached 33.5% and specificity recovered to 72.4%. The study shows why physical measurements can matter when computational scores mistake an assay artifact for a useful molecule.
VERID tracks a lab robot's physical state
A separate report introduces VERID, a persistent digital model of a laboratory workcell. It tracks the robot deck, containers, liquid volumes and pipette tips; simulates a proposed protocol before dispatch; and keeps synchronized action logs and video. In the reported tests, agents caught impossible or resource-invalid actions and resumed work after equipment or communication failures.
Those VERID demonstrations used colored dye and water, not biological samples. The tests examined whether the controller could track volumes, detect geometric conflicts and help diagnose fluid-handling errors. They are a foundation for later biological experiments, not proof that an autonomous wet lab can safely run complex research on its own.
Protein binders remain research leads
In the third study, researchers combined local target-region cropping with agent-guided search to design binders for fibroblast activation protein alpha (FAP), a 719-residue protein found in solid tumors. Cropping around a proposed binding site reduced GPU use by 88% in the reported large-target workflow. Under a fixed compute budget, a human-agent hybrid raised the computational hit rate from 5.06% to 9.33% while using 42% fewer GPU hours than a human-only workflow. Here, a computational hit meant a design passed the paper's fixed filters for predicted interface quality and structural agreement: ipSAEmax above 0.45 and interface-core C-alpha RMSD below 3 angstroms. It was a model-screening result, not an experimentally confirmed binder.
Researchers then tested selected designs with yeast-surface display. The apparent equilibrium dissociation constant (KD) summarizes binding strength; lower values indicate tighter binding. Four enriched clones had apparent KD values of 3.0 to 9.1 nM in binding assays. The authors say further work is needed to test soluble expression, binding specificity and whether the designs inhibit FAP activity. A favorable affinity measurement is one early filter; it does not establish drug-like behavior, safety or clinical benefit.
The three reports also show different degrees of autonomy. Humans performed the aptamer wet-lab procedures, and the strongest FAP computational result came from a hybrid in which a researcher set the overall strategy and an agent optimized parts of the search. Google says its test systems were chosen to avoid infectious or gain-of-function risks and that people reviewed designs and protocols. The next test is whether these pieces can be joined into reproducible, human-directed cycles that discover useful molecules without confusing model scores, assay artifacts and real biological activity.
Image: Nick Berndt operating an automated liquid-handling system used for eDNA sample processing. Photo by Scott Covington, U.S. Fish and Wildlife Service, public domain, via Wikimedia Commons. The image is illustrative and is not from Google's experiments. Image details.
.jpg){kind=link}
Sources
- Google Research: “Grounding AI agents in biological reality” (October 16, 2026)
- Delley et al., “Semantic Representation for Agentic Self-Play in Automated Laboratories”
- Cutts et al., “Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery”
- Kim et al., “Agent-guided epitope excision enables economical de novo binder design”