AI Agents Improve DNA Aptamers in 3-Round Wet-Lab Loop

AI Agents Learned From Wet-Lab Results to Design Better DNA Binders

TL;DR: Google Research ran AI agents through three rounds of DNA aptamer experiments against two proteins, and bead-corrected hit rates climbed to 42.1% for α-thrombin and 33.5% for myoglobin after agents revised designs from lab feedback.

AI molecular design has largely worked on computers, with laboratory testing saved for a final check rather than used to steer the next design. A team from Google Research, Stanford University, Google Cloud and Google DeepMind tested whether large language model reasoning agents could improve by working in repeated laboratory cycles, according to the Google Research publication titled “Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery.” The authors selected protein-binding DNA aptamers as a fast experimental system, described as supporting 1–2 week validation cycles at $1–$3 per candidate.

The project tasked agentic pipelines with creating aptamer binders for α-thrombin and myoglobin through three design-test-improve cycles, with a budget of about 1,000 candidates per round and about 4,400 total sequenced constructs. Hit rates reported as statistically significant and bead-corrected among newly designed candidates increased from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3 for α-thrombin, and from 10.2% in Round 1 to 3.7% in Round 2 and 33.5% in Round 3 for myoglobin. The same publication reports that single-clone titrations for eight α-thrombin candidates matched pooled sequencing enrichment with Pearson r=0.90, with literature-derived motifs reaching apparent KD=24.6–133 nM and strictly de novo scaffolds showing KD>1 μM.

Why DNA aptamers were selected as the test system

The researchers describe aptamer discovery as work commonly used in clinical biomarker discovery efforts, according to the Google Research blog post titled “Grounding AI agents in biological reality: Building blocks for autonomous wet labs.” That post, dated October 16, 2026 and authored by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, states that a complete aptamer design-test-analyze cycle takes 7–10 days.

The speed comes from chemistry and readout choices. DNA can be chemically synthesized in pooled libraries and read out directly by sequencing, the blog post states. In each round, agents design candidate sequences, those sequences are screened against target proteins on magnetic beads to isolate functional binders, and sequencing returns quantitative data for the next cycle.

The preprint identified as arXiv:2610.24806 (2026) lists authors Zachary Cutts, John Coller, Rachel Chen, Hua Tang, Mathias Bellaiche, Ryder Easterlin, Dina Zverinski, Petar Sirkovic, Avinatan Hassidim, Yossi Matias, Pooja Rao and Annalisa Pawlosky. Affiliations listed are Google Research, Department of Genetics at Stanford University in Stanford, CA, USA, Google Cloud and Google DeepMind. The author note states work was performed while H.T. was a Visiting Faculty Researcher at Google Research and identifies equal leadership by co-senior authors.

What “largely in silico” means for current pipelines

The publication states that despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for final validation rather than iterative improvement. It frames as unproven whether large language model-based reasoning agents can meaningfully accelerate de novo discovery by iterating with real-world lab data.

The blog post makes a related point that computational models do not capture the full picture of what happens in a wet-lab reaction and cannot replace empirical validation. It says agents need real experimental readouts to refine hypotheses. The publication adds that computational predictors sit on a spectrum of reliability tied to training data and cannot replace wet-lab validation for any modality.

The publication also describes agent operation more broadly. It says AI agents operate within software harnesses that interleave reasoning with tool execution, code generation, literature retrieval and persistent memory, allowing decisions to be updated as new data arrive. It describes multi-agent systems for scientific discovery, often termed “AI co-scientists,” as systems that have begun to accelerate biological research by generating testable hypotheses, mining sequence databases and designing candidate molecules.

How the three design-test-improve cycles were organized

The aptamer sandbox used two protein targets, α-thrombin and myoglobin, across three rounds. The budgeted scale was about 1,000 candidates per round, totaling about 4,400 sequenced constructs counted in the publication.

According to a TokenPost report titled “Google Details Three AI Projects Aimed at Autonomous Wet Labs,” the design-to-screen cycle lasted about 10 days, with pooled DNA synthesis and reagents estimated at $3 per candidate. That report breaks out new-candidate counts that align with the published percentages.

For α-thrombin, the TokenPost account lists no hits among 167 new candidates in round one, 54 of 208 in round two and 131 of 311 in round three. For strictly de novo candidates without known thrombin motifs, it lists 106 of 286 in the third round. For myoglobin, it lists 20 of 196 in round one, nine of 244 in round two and 184 of 549 in round three. Those counts correspond to the bead-corrected rates in the research publication.

α-Thrombin results rose across all three rounds

For α-thrombin, statistically significant, bead-corrected on-target hit rates among newly designed candidates rose from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3, according to the research publication. Excluding literature motif descendants, the Round 3 rate was 37.1%.

The TokenPost breakdown expresses the same outcome as counts: 0 of 167, then 54 of 208, then 131 of 311. The strictly de novo subset in Round 3 was 106 of 286 at 37.1%. The publication presents the 37.1% figure as the rate when descendants of literature motifs are removed.

Follow-up testing used orthogonal single-clone bead-binding titrations of eight full-length α-thrombin candidates. Those measurements confirmed quantitative concordance with pooled sequencing enrichment at Pearson r=0.90. In those assays, literature-derived motifs engineered by the agents to reduce off-target reactivity in the pooled screen reached preliminary apparent affinities of KD=24.6–133 nM, while strictly de novo scaffolds showed low-micromolar binding at KD>1 μM.

Myoglobin results dipped, then recovered

Myoglobin followed a different trajectory. Bead-corrected hit rates were 10.2% in Round 1, 3.7% in Round 2 and 33.5% in Round 3 among newly designed candidates, according to the research publication. The TokenPost account gives the underlying counts as 20 of 196, nine of 244 and 184 of 549.

The publication links the middle-round pattern to use of myoglobin binding data to exploit the high-affinity G-quadruplex binder space. It then reports diagnosis of cross-reactivity as on-target specificity dropped from 75.0% to 19.1%. The pipelines subsequently recovered by developing new strategies that restored myoglobin specificity to 72.4%.

The blog post summary describes agents improving designs across cycles from the same binding data. The publication states that empirical lab data proved more valuable than existing computational proxies in this setting. It says iterative experimental feedback enabled AI agents to diagnose assay artifacts and refine molecular designs across rounds.

What the G-quadruplex detour revealed

The myoglobin arm showed how a high-affinity structural class can create misleading progress. The publication reports initial exploitation of the G-quadruplex binder space from myoglobin binding data. Specificity then fell from 75.0% to 19.1%, indicating cross-reactivity.

The agents then diagnosed cross-reactivity and changed strategy. Specificity returned to 72.4% in the later work, while the bead-corrected on-target hit rate reached 33.5% in Round 3. The research publication presents this sequence as exploitation, diagnosis and recovery within the same closed loop.

The broader point made by the authors is that experimental feedback exposed a misleading optimization target. A summary of the work notes that this offers a transferable approach for distinguishing genuine biological improvements from artifacts in automated discovery workflows. The publication itself says the result demonstrates how iterative experimental feedback lets AI agents diagnose assay artifacts and refine designs.

How binding was measured and corrected

The screening approach exposed candidate pools to target proteins on magnetic beads, isolated functional binders and sequenced them to produce quantitative data for the next cycle, according to the blog post. The publication describes reported hit rates as bead-corrected and on-target, and as statistically significant among newly designed candidates.

Control handling mattered because beads themselves can retain sequences. The bead correction is presented as the adjustment used to separate target-associated enrichment from bead-associated background. Only after that correction do the reported Round 1 to Round 3 changes constitute the headline hit-rate progression.

The eight-candidate single-clone work provided an independent check. Those bead-binding titrations were performed on full-length α-thrombin candidates outside the pooled format. Their agreement with pooled sequencing enrichment at Pearson r=0.90 is presented as quantitative concordance between the high-throughput screen and lower-throughput titrations.

VERID: the separate robot-control building block

The October 16, 2026 blog post presents the aptamer loop alongside a separate robotics effort called VERID, short for VERified Infrastructure for Discovery Laboratories. VERID is described as a semantic digital twin of the laboratory workcell, namely a live, stateful 3D volumetric entity tree that tracks everything from robotic arm geometry to individual well volumes, reagent identities and pipette tip inventory.

VERID is built on top of the open-source PyLabRobot framework. It provides live 3D state tracking, fast virtual dry runs, hardware safety checks and synchronized video and action logs, according to the blog post. Agents can check deck layout, well volumes and available tips before acting, simulate protocols in seconds to catch collisions or empty wells before touching reagents, and have invalid commands blocked before they reach the robot.

The case studies used dye and deionized-water mixtures, not aptamers or therapeutic candidates. In one example, synchronized video and execution traces caught uneven dye stratification caused by dispensing too close to the well floor, allowing the agent to refine mixing. In the same campaign, persistent digital twin state allowed recovery mid-run from a plate-reader failure without re-pipetting. In another case, the simulator rejected an 8-channel aspiration that would not fit in a narrow trough, after which the agent replanned and designed a targeted factorial diagnostic isolating bias from residual tip-wall carryover. In a closed-loop self-play campaign, an agent iteratively designed and ran robot experiments to reconstruct blinded dye mixtures from overhead camera images, improving accuracy more than fivefold.

What remains separate and what comes next

Google’s blog post authors state a long-term vision in which AI agents will not only ideate and analyze, but also run lab experiments to test hypotheses, learn from results and self-improve, all driven by human scientific agency. The same post identifies two intertwined challenges: frontier models still do not reason reliably about three-dimensional geometry and physical state, and real-world wet-lab experiments have not kept pace with in-silico design, making experimental readouts a bottleneck.

The robot-control tests, aptamer screening and protein-binder design were separate experiments, according to the TokenPost report. VERID’s robot-control system and closed-loop protein discovery are planned for connection, but that integration remains a next step. Safeguards described in project summaries include sequence screening, simulation before robot execution, and human authorization and supervision.

The aptamer result does not establish a therapy or a deployed design service. It demonstrates a wet lab in a closed loop with AI systems at the stated scale of about 1,000 candidates per round and about 4,400 sequenced constructs, with the reported affinity split between engineered literature-derived motifs at 24.6–133 nM apparent KD and strictly de novo scaffolds above 1 μM. The authors’ stated conclusion is that empirical lab data was much more valuable than existing computational proxies for guiding agents across rounds.