DNA aptamers are folded strands that bind selected molecules, useful for detecting biological markers. In research announced on 16 October, Google’s AI agents designed these recognisers, received laboratory measurements, then tried again. By the third round, 33.5% of their new candidates met the study's criteria for binding the protein myoglobin. The path there included a sharp reversal.[2]
The test put target proteins on magnetic beads. Candidate DNA strands were exposed to them, and sequencing revealed which strands had been retained. The measurements then informed the next designs. Google describes a cycle lasting roughly a week to ten days, with humans authorising designs and supervising the laboratory work.[2]
Myoglobin: the new designs that passed the binding screen
| Round | Passing candidates / new candidates | Hit rate |
|---|---|---|
| 1 | 20 / 196 | 10.2% |
| 2 | 9 / 244 | 3.7% |
| 3 | 184 / 549 | 33.5% |
In round two, the agents favoured G-quadruplexes, a family of folded DNA structures. Raw binding rose, but many strands also stuck to wrong proteins or bare beads. The agents then limited promiscuous motifs and explored alternatives. Among significant binders in the multi-protein test, the share binding only to myoglobin recovered from 19.1% to 72.4%.[2]
A bead is part of the apparatus, not the molecular target. A design that clings to it can prosper in the initial readout while failing the scientific task. The recovery therefore depended on changing which apparent successes deserved to guide the next search.
What the molecules could actually do
For the other target, α-thrombin, the team separately tested eight individual candidates. Four adaptations of known motifs had preliminary apparent dissociation constants of 24.6–133 nanomolar; two entirely new scaffolds showed binding onset at one micromolar, while two remained near background. Lower dissociation constants indicate tighter binding. The estimates remained preliminary because the tested concentration range did not establish a full saturation plateau.[1]
The 12 October manuscript also says the agent setup changed between rounds and lacked matched no-feedback controls. The campaign demonstrates an improving experimental workflow; it cannot assign the whole gain to agent reasoning alone.[1]
For molecular discovery, even an imperfect first binder can be a useful starting point if its behaviour is measured clearly. Here, the next promising idea emerged from a failed interpretation of the previous one. The laboratory gave the search a way to learn what its molecules were actually holding on to.
Sources & further reading
- Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery Google Research and collaborators · 12 Oct 2026 · Accessed 18 Oct 2026
- Grounding AI agents in biological reality: Building blocks for autonomous wet labs Google Research · 16 Oct 2026 · Accessed 18 Oct 2026