AI BioDesign: How Self-Driving Labs Close the Loop on Discovery

The phrase "self-driving lab" still sounds like science fiction, but the machinery is already running. Across drug discovery, synthetic biology, and materials science, a new class of system is beginning to do what software did to information: design experiments, run them with robots, read the results, and rewrite its own next move. The loop is the point.

From pipelines to loops

Classical R&D is linear. A scientist forms a hypothesis, designs an assay, runs it by hand, waits, and interprets. An AI BioDesign stack inverts this. Machine-learning models propose candidates — a mutated enzyme, a synthesis recipe, a drug-like molecule — and a layer of lab automation executes them: liquid handlers, plate readers, incubators, robotic arms. Measurements stream back into the model, which proposes the next batch. Each cycle sharpens the predictions.

The concept has a name in the literature: self-driving laboratories (SDLs), and Nature flagged the category among the technologies to watch in 2025. A 2025 Nature Communications manifesto by Canty, Bennett, Brown and colleagues frames SDLs as the fusion of automated experimentation, AI, and real-time analytics — a closed loop that can explore chemical, biological, and physical parameter spaces in compressed time frames.

Real systems, real output

The most cited demonstration in materials is the A-Lab at Lawrence Berkeley National Laboratory. Reported in Nature in 2023, the A-Lab combined ab-initio phase-stability data from the Materials Project and Google DeepMind, machine-learning models trained on roughly 30,000 published synthesis procedures, and a robotic line that dosed, heated, and characterized powders with X-ray diffraction. Over 17 days of continuous operation it synthesized 41 of 58 target inorganic compounds — work that would have taken human chemists months. DeepMind's companion model, GNoME, had independently proposed more than 2.2 million stable crystal structures.

In biology, the University of Illinois iBioFoundry offers a cleaner illustration of the loop. Led by Huimin Zhao, the team (reported in Nature Communications in 2025, DOI: 10.1038/s41467-025-61209-y) coupled an AI model that predicts beneficial enzyme mutations with the foundry's automated protein-building and testing machines. Starting from a known enzyme, the AI proposes sequence changes; robots build and characterize them; the data retrains the model. The system lifted the activity of a feed-additive enzyme 26-fold and a synthesis catalyst 16-fold with 90-fold better substrate preference. "It's a step toward a self-driving lab," said iBioFoundry manager Stephan Lane.

Japan's Autonomous Lab (ANL), described in Scientific Reports in 2025, takes a modular, Bayesian-optimization approach to optimize culture media for engineered E. coli — evidence the model is spreading beyond elite institutions.

The commercial stack

Two business models have crystallized. Cloud labs — Emerald Cloud Lab (founded 2012) and Strateos (formerly Transcriptic, founded 2014) — let any scientist script experiments and run them on shared robotic infrastructure via API. Emerald's platform treats instruments as programmable functions in a symbolic language derived from Wolfram, and its partnership with Carnegie Mellon produced the first university-based autonomous cloud lab in 2024, capable of running more than 100 experiments simultaneously. Strateos has integrated with partners including Eli Lilly and Bayer for remote, closed-loop biology workflows.

The foundry model is different: Ginkgo Bioworks, founded in 2008 out of MIT, built physical "foundries" where robots prototype thousands of organism variants, running the design-build-test-learn cycle as a service for flavors, fragrances, agriculture, and pharma. Its codebase of engineered biological parts feeds back into each new design.

At the deepest end sits Recursion, a "TechBio" company running more than two million experiments per week across automated labs in Salt Lake City and Oxford. Its phenomics platform images cells at massive scale — millions of high-content microscopy images feeding computer-vision and foundation models such as Phenom-2 (1.9 billion parameters, trained on 8 billion images) and MolGPS (3 billion parameters). Recursion's BioHive-2 supercomputer, built on NVIDIA H100 GPUs, trains these models, and partnerships with Roche/Genentech, Sanofi, Bayer, and Merck KGaA turn the data into drug programs. In late 2024 it acquired Exscientia, folding precision chemistry design into the same loop.

Isomorphic Labs, the DeepMind spinout, represents the "in silico" pole: AlphaFold and its 2024 successor AlphaFold 3 model molecular interactions, and the company signed roughly $3 billion in partnered deals with Novartis and Eli Lilly in 2024 before raising $600 million in 2025 — now preparing its first AI-designed molecules for human trials in oncology and immunology.

Why this is a strategic frontier

ORIGINAL INSIGHT: AI's first decade of economic impact lived in software — text, images, code — domains where the cost of a wrong answer is a bad sentence, not a spilled culture. Biology is the opposite: the design space is effectively infinite (the Illinois team notes a typical enzyme has more possible variants than atoms in the universe), the experiments are slow and expensive, and the payoff — new medicines, materials, and sustainable chemistry — is enormous. Self-driving labs are where AI stops being a copilot for documents and becomes the control system for physical discovery itself. The strategic prize is not any single product but ownership of the loop: whoever controls the tightest design-make-test-learn cycle accumulates proprietary data fastest, and data is the moat. This is why cloud labs, foundries, and pharma are racing to close the loop now, while the field is still young.

Honest limits

The hype needs ballast. Most SDLs still require human-defined assays and goals; the "wet lab" — culturing, contamination control, sample handoffs — resists full automation, and A-Lab's failures were often plain experimental difficulties a person could fix in minutes. Integration is the bottleneck: tying dissimilar instruments into one orchestrated loop is harder than training the model. McKinsey has estimated that lab automation could bring medicines to market about 500 days faster and cut development costs by roughly 25 percent, but those gains accrue only when the loop runs end-to-end. For now, self-driving labs are powerful accelerators, not autonomous scientists — and human oversight remains both an ethical and a practical necessity.