AI-Run 'Self-Driving Labs' Are Automating Protein and Enzyme Discovery

AI-Run 'Self-Driving Labs' Are Automating Protein and Enzyme Discovery

For most of its history, protein engineering has been a grind of human intuition and repetitive labor: a scientist guesses at a mutation, spends days at the bench building and testing it, squints at the result, and guesses again. In 2026, that loop is increasingly being closed by machines that never sleep. A new class of "self-driving labs" — AI-powered biofoundries that couple machine-learning design with robotic build-and-test hardware — is now running the entire design-build-test-learn (DBTL) cycle with little or no human hands on the pipette.

The clearest recent proof point came out of the University of Illinois Urbana-Champaign, where a team led by chemical and biomolecular engineering professor Huimin Zhao built what it calls a generalized platform for autonomous enzyme engineering. Reported in Nature Communications in 2025, the system runs on the university's iBioFoundry: an AI algorithm proposes sequence changes, robots synthesize and assemble the new enzyme variants, automated assays measure how well they work, and the results feed straight back into the model to shape the next round — no scientist required in the loop.

The numbers are the point. Working over four autonomous rounds spanning about four weeks, and building fewer than 500 variants per target, the platform improved an Arabidopsis thaliana halide methyltransferase by 90-fold for substrate preference and 16-fold for ethyltransferase activity, and boosted a Yersinia mollaretii phytase — an enzyme added to animal feed — by 26-fold in activity at neutral pH.

"Enzymes have been increasingly used in energy production, in therapeutics, even in consumer products like laundry detergent," Zhao said. "But they are not as widely used as they could be, because they still have limitations. Our technology can help address those limitations efficiently."

The Illinois work builds on a broader movement. At the University of Wisconsin–Madison, biochemist Philip Romero and colleague Jacob Rapp developed SAMPLE — Self-driving Autonomous Machines for Protein Landscape Exploration — a platform driven by an intelligent agent that learns sequence-function relationships, designs proteins, and dispatches them to a fully automated robotic system for testing. In a study published in Nature Chemical Engineering, four SAMPLE agents were set loose to make glycoside hydrolase enzymes more heat-tolerant. Despite quirks in how each agent searched, all four converged on variants at least 12 degrees Celsius more stable than the starting sequences.

The appeal is as much about speed and stamina as accuracy. A single round of testing that takes a human researcher three to four days can be done by the self-driving system in roughly nine hours — and the machine immediately starts the next round, working around the clock. "We set and forget it," Romero has said of the platform, capturing the shift from bench work to supervision.

The frontier is now moving toward systems that non-experts can actually operate. A February 2026 preprint on bioRxiv described an "AI-native" autonomous biofoundry built on a cloud-edge architecture, with an agent-native control layer powered by large language models and Anthropic's Model Context Protocol. The pitch is that a researcher can orchestrate an entire DBTL campaign through natural language rather than custom scripts. Under the hood it stitches together deep phylogenetic mining, the zero-shot protein language model ESM-2, and supervised active learning to navigate rugged fitness landscapes. As a proof of concept, the team evolved a Family B DNA polymerase for CoolMPS sequencing chemistry, reporting a hit rate above 66 percent and variants with a 37 percent reduction in sequencing error rate — in just three autonomous rounds.

Two 2025 review articles in Current Opinion in Biotechnology frame these systems as a genuine 2026 trend rather than a set of one-off demos, arguing that AI-powered biofoundries are shifting both protein engineering and metabolic engineering from labor-intensive manual work toward continuous autonomous experimentation — while also flagging that the field still lacks the transparent, reproducible, and accessible standards needed for results to be trusted and shared.

Why It Matters

The strategic significance of self-driving labs is that they close the design-build-test-learn loop and put it under machine control. Each turn of that loop generates clean, structured experimental data, which trains better models, which propose better experiments — a compounding flywheel that manual research cannot match. For enzymes, that means faster paths to industrial biocatalysts for greener manufacturing, feed additives, and biofuels. For drug discovery, the same closed-loop machinery can be pointed at antibodies, therapeutic proteins, and the enzymes behind next-generation DNA sequencing and synthesis. And because the platforms are increasingly "generalized," the same hardware can be retargeted from one protein family to another without rebuilding the lab. Perhaps most consequentially, by stripping out the requirement for deep specialist expertise, these systems threaten to democratize a discipline that has long been gated by a handful of elite labs — while raising familiar questions about reproducibility, biosecurity, and who audits an experiment that no human ran.

What to Watch

Watch for the shift from academic proofs-of-concept to commercial cloud biofoundries where scientists rent autonomous DBTL cycles by the run. Watch whether LLM-and-MCP "agentic" control layers, like the one in the February 2026 preprint, prove robust enough to hand real decision-making to language models rather than narrow optimizers. Watch for standardization efforts — shared data formats, benchmark proteins, and reproducibility protocols — that the review literature says the field urgently needs. And watch the metrics that actually matter for adoption: not just fold-improvements in a demo enzyme, but cost per validated variant, cycle time, and whether these platforms can deliver a molecule that ends up in a factory, a feed trough, or a clinic.