A research team reports steps toward integrating in‑silico AI agents with real wet‑lab experimentation by developing a semantic digital twin (VERID), running closed‑loop aptamer discovery cycles, and accelerating de‑novo protein binder design. The presented systems let agents plan and validate experiments against a live 3D lab model, iterate on experimental readouts to refine designs, and reduce computational costs for protein folding to speed cycles. Safety layers and human oversight are emphasized throughout.
The problem
Two linked challenges constrain autonomous, agent‑driven laboratory discovery. First, frontier models still lack reliable reasoning about 3‑dimensional geometry and physical states, which is critical for automating liquid‑handling robots. Second, although agentic hypothesis generation and computational design advance rapidly, real wet‑lab experimental throughput and feedback have not matched that pace, making experimental readouts a bottleneck.
VERID: grounding agents in a digital twin
VERID (VERified Infrastructure for Discovery Laboratories) is a semantic digital twin that models a laboratory workcell as a live, stateful 3D volumetric entity tree. It tracks elements ranging from robotic arm geometry to individual well volumes, reagent identities, and pipette tip inventories.
Built on the open‑source PyLabRobot framework, VERID provides four core capabilities:
- Live 3D state tracking: agents can check deck layout, well volumes, and available tips before acting.
- Fast virtual dry runs: protocols are simulated in seconds to catch collisions or empty wells before touching reagents.
- Hardware safety checks: invalid commands are blocked before reaching the robot.
- Synchronized video and action logs: a complete record of every run for rigor, reproducibility and debugging.
Case studies illustrate these benefits. Synchronized video and execution traces revealed uneven dye stratification caused by dispensing too close to the well floor, allowing an agent to refine mixing strategy. The persistent digital twin state enabled recovery mid‑run from a plate‑reader failure without re‑pipetting. When the simulator rejected a physically impossible 8‑channel aspiration in a narrow trough, the agent replanned and designed a factorial diagnostic that uncovered subtle 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 images, identifying optical and fluidic artifacts across runs and improving accuracy more than fivefold.
Comprehensive execution and trace logs — including dwell times, evaporation, and plate layout — provide the data foundation for agents to separate biological signal from experimental noise while exploring many candidates in parallel.
Learning with aptamers in a lab‑in‑the‑loop
The paper “Aptamers as a sandbox for lab‑in‑the‑loop agentic molecular discovery” uses DNA aptamers that bind proteins to prototype closed‑loop agentic discovery across three design‑test‑iterate cycles. Aptamer cycles are relatively fast (7–10 days) because DNA can be chemically synthesized in pooled libraries and read out by sequencing.
In each round agents design candidate sequences, screen them against target proteins on magnetic beads to isolate binders, then sequence survivors to return quantitative data for the next cycle. Across three rounds on two targets (myoglobin and α‑thrombin), AI‑driven rational and de‑novo design raised confirmed on‑target hit rates to 33.5% for myoglobin and 42.1% for α‑thrombin, although each target required identifying and overcoming different wet‑lab failure modes.
Specific observations:
- α‑thrombin: Round 1 produced 0% binders because fixed primer overhangs disrupted folding in 48% of sequences. After agents introduced clamped flanking stems in Round 2 and diversified in Round 3, hit rates rose to 26.0% and then 42.1% at 98.5% specificity.
- Myoglobin: In Round 2 agents prioritized raw affinity and heavily enriched G‑quadruplex motifs (63% of designs). While pull‑down increased, lab data showed most G‑quadruplexes were non‑specific stickier binders to off‑targets and bare beads, causing specificity to fall from 75.0% to 19.1% and bead‑corrected hit rate to drop from 10.2% to 3.7%. Recognizing this artifact from Round 2, agents autonomously revised design rules in Round 3: capping promiscuous motifs, subtracting bead binders, and shifting to non‑G4 structural scaffolds. These changes recovered specificity to 72.4% and increased on‑target hit rate ninefold to 33.5%.
The study also analyzes how agents reason about data and protocols like scientists, while occasionally exploring sequence spaces that structural biologists might not prioritize.
Accelerating de‑novo protein binder design
Proteins are harder to close the loop on because they require biological expression; each iteration adds cellular or cell‑free selection, expression, and folding steps, and the sequence space is larger. This increases bottlenecks both computationally and experimentally.
In “Agent‑guided epitope excision enables economical de‑novo binder design,” the workflow was applied to fibroblast activation protein alpha (FAP), a large, challenging receptor overexpressed in solid tumors. In a few experimental runs the top designs yielded binders with single‑digit nanomolar affinity, a level of tight, drug‑like binding typically reached only after extensive lab optimization.
A key computational improvement was to excise only the local binding pocket during folding, which reduced the re‑folding cost — the GPU‑dominant step that scales quadratically with target length — and accelerated each design cycle more than eightfold without losing structural accuracy. This makes tactical choices (where to bind and how much surrounding structure to include) parameters that language‑model agents can optimize.
In a head‑to‑head comparison at equal compute, human–agent collaboration — a scientist setting high‑level strategy and an agent executing tactics — nearly doubled the in‑silico hit rate of the scientist alone and found hits across a broader range of target sites. To scale wet‑lab screening, the team operationalized yeast surface display with fluorescence‑activated cell sorting (FACS) to screen large libraries of engineered variants in parallel by fluorescence.
The next step is to integrate VERID’s traceable automation with closed‑loop learning for protein discovery so agents can directly navigate complex biological search spaces in the wet lab.
Safety
Following Google’s Frontier Safety Framework and a bioresilience approach, the project builds safeguards across computational and wet‑lab workflows:
- Safe testbeds and screening: deliberately chosen, well‑controlled non‑infectious systems without gain‑of‑function risk; every designed sequence undergoes biosecurity screening.
- Simulation before execution: through VERID, agents never send unverified code directly to liquid‑handling robots; protocols are dry‑run and validated against the digital twin first.
- Human oversight: people authorize and review designed sequences, verify wet‑lab protocols and supervise robot runs.
As agents assume more of the design–test cycle, the authors note open problems that must be addressed with opportunities for human agency: ensuring agents optimize for true target binding rather than experimental shortcuts, verifying multi‑step wet‑lab protocols, and continuously monitoring agent reasoning and execution traces.
Conclusions
Combining a live digital twin, rapid closed‑loop aptamer cycles, and accelerated de‑novo protein design forms foundational components for human‑directed autonomous discovery in the life sciences. The presented results — including reproducible automation, iterative aptamer improvements, and single‑digit nanomolar bind‑ ers for a difficult target — mark progress toward agent‑driven wet labs under layered safety and human supervision.
Acknowledgements
The research is led by Annalisa Pawlosky with Pooja Rao and Petar Sirkovic. Co‑authors and collaborators include Zach Cutts, John Coller, Rachel Chen, Mathias Bellaiche, Hua Tang, Ryder Easterlin and Dina Zverinski (aptamer discovery); Jun Kim, Artiom Myaskovsky, Oscar Donosa, Jamie Rogers and Lawrence Chan (protein design); and Yves Delley, John Coller, Zach Cutts, Alexander Daryin, Mathias Bellaiche, Jun Kim, Oscar Donosa, Srinidhi Seshadri, Tanner Clary and Sandy Klemm, with technical and experimental contributions from Xinyi Deng, Sanath Ramachandran and Jiawu Bi (VERID automation). The team thanks Reuven Sayag, Avinatan Hassidim, Ronit Levavi Morad, Yossi Matias and James Manyika for leadership guidance, and Matthias Heiler, Burak Gokturk, Bill Jia and Pushmeet Kohli for support, along with reviewers and contributors listed in the acknowledgements.
