Google Research has published “Grounding AI agents in biological reality: Building blocks for autonomous wet labs” (16 October 2026), by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky. The post describes AI agents planning and running protocols on liquid-handling robots, closed-loop DNA aptamer design, and early protein binder work against a cancer-associated target.
Every result below is Google’s own, drawn from three self-published reports linked from the blog (VERID, aptamer sandbox, epitope excision PDFs on gstatic). They are not peer-reviewed. This is lab prototype research, not a shipped product, and the binders are early-stage research molecules, not drugs.
VERID: a lab digital twin on PyLabRobot
Google calls its framework VERID — “VERified Infrastructure for Discovery Laboratories” — “a semantic digital twin of the laboratory workcell, i.e., a live, stateful 3D volumetric entity tree” that tracks robotic geometry, well volumes, reagent identities and pipette-tip inventory. It is “Built on top of the open-source PyLabRobot framework.”
VERID’s four stated capabilities:
- Live 3D state tracking — check deck layout, well volumes and available tips before acting
- Fast virtual dry runs — simulate protocols in seconds to catch collisions or empty wells
- Hardware safety checks — “Block invalid commands before they reach the robot”
- Synchronized video and action logs — full records for reproducibility and debugging
In a closed-loop self-play campaign, Google writes, an agent reconstructed blinded dye mixtures from overhead camera images, “identifying its own optical and fluidic artifacts across runs to improve accuracy more than fivefold.” Other case studies on the blog describe catching dye stratification, recovering from a plate-reader failure mid-run, and replanning when an 8-channel aspiration would not fit a narrow trough.
Aptamers: closed-loop design in 7–10 days
In the companion report framed as an aptamer “sandbox,” Google says “A complete aptamer design-test-analyze cycle takes just 7–10 days.” Agents design sequences; libraries are screened on magnetic beads; sequencing feeds the next round.
Over three rounds on myoglobin and α-thrombin, confirmed on-target hit rates rose “to 33.5% and 42.1% for myoglobin and α-thrombin, respectively.”
On α-thrombin, Round 1 “yielded 0% binders” because fixed primer overhangs “disrupted folding in 48% of sequences.” After clamped flanking stems and diversification, hit rates climbed to 26.0% and then 42.1% “at 98.5% specificity.”
On myoglobin, Round 2 G-quadruplex “sticky” binders cut specificity “from 75.0% to 19.1%” (bead-corrected hit rate from 10.2% to 3.7%). Agents revised design rules in Round 3, recovering specificity (72.4%) and a “ninefold” rise in on-target hit rate (33.5%).
Protein binders: nanomolar FAP designs
For de novo protein binders, Google applied the workflow to fibroblast activation protein alpha (FAP), described as “a large, challenging receptor overexpressed in solid tumors.” “In just a few experimental runs, the top designs yielded binders with single-digit nanomolar affinity.”
Epitope excision — folding only the local binding pocket — “accelerated each cycle more than eightfold” (the blog figure labels structure prediction 8.6x). At equal compute, human–agent collaboration, “where a scientist sets the high-level strategy and an agent drives tactical execution,” “nearly doubled the in-silico hit rate of a scientist working alone.” Wet-lab screening used yeast surface display with FACS.
Again: these are Google’s reported affinities and speed-ups from self-published work. They are not clinical candidates and not an approved therapy.
Safety measures Google states
The blog’s safety section, “in line with Google’s Frontier Safety Framework and our approach to bioresilience,” lists:
- Safe testbeds and screening: “We deliberately selected well-controlled, non-infectious biological systems without gain-of-function risk” as testbeds; “Every designed sequence goes through biosecurity screening.”
- Simulation before execution: “Through VERID, agents never send unverified code directly to the liquid handling robots. Every protocol is dry-run in simulation and validated against the digital twin before touching reagents.”
- Human oversight: “Humans directly authorize and review designed sequences, verify wet-lab protocols, and supervise robot runs.”
Google also notes open problems as agents take on more of the design-test cycle — including agents optimising for true binding rather than experimental shortcuts, and continuous monitoring of reasoning and execution traces.
Why it matters
TSN has covered adjacent lab-AI threads: DeepMind’s Antigravity antimicrobial peptide candidate (still needing lab testing), OpenAI’s Rosalind academic plan for university labs, and DeepMind’s Invisible Signature watermarking for AI-designed proteins. Google’s new post is useful as a documented closed-loop stack — digital twin, aptamer cycles, binder design — with explicit human-in-the-loop and biosecurity language. It is still a research blog + PDF trio, not a peer-reviewed package or a product launch.
What’s still unknown
- Independent replication of the fivefold dye, aptamer hit-rate and nanomolar FAP claims
- Peer-review outcomes for the three supporting PDFs
- How far VERID generalises beyond Google’s own workcells
- Any timeline for external access to the tooling (none stated)
What this means
Google Research says AI agents, grounded through VERID on PyLabRobot, can dry-run and run wet-lab protocols, close the loop on aptamers in 7–10 days, and produce single-digit nanomolar FAP binders in a few runs — while humans authorise sequences, protocols are simulated first, and testbeds stay non-infectious without gain-of-function risk. Treat every percentage and affinity as Google’s self-published claim, not a peer-reviewed result or a medicine.
Sources
- [1] Google Research: Grounding AI agents in biological reality: Building blocks for autonomous wet labs, 16 October 2026
- [2] VERID semantic lab report (PDF)
- [3] Agentic aptamer sandbox report (PDF)
- [4] Epitope excision / binder design report (PDF)
- [5] PyLabRobot documentation
Related stories
- Google DeepMind Shows Its Antigravity Agent Finding a New Antimicrobial Peptide Candidate, Which Still Needs Lab Testing
- OpenAI Introduces the Rosalind Academic Plan for US University Labs
- The Invisible Signature: Google DeepMind Has Started Watermarking AI-Designed Proteins
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
