Lab Newsletter — July 20, 2026: Design, Build, Test — and Safeguard?

Three items today trace one arc: AI can now design biology’s catalysts, machines can build and test them, and — the part that hasn’t kept pace — almost none of it ships with guardrails.

🧫 Generative AI can finally design enzymes

A Feb 2026 review (Middendorf & Ferruz) marks a real turning point: “after more than one decade of low success rates for computationally designed enzymes, generative AI models are now frequently used for designing proficient enzymes” — mature enough, they argue, for industrial use. The shift is from structure-first to function-first design, with generative models (ProGen, ESM-2, ProteinGAN, diffusion) producing “synzymes” that catalyze reactions nature never evolved. Why it matters for the lab: this is generative modeling graduating from describing biology to designing it — the same generative-imaging thread our ProtiCelli work pulls on, now producing functional molecules.

🤖 Autonomous biofoundries close the loop

Design is only half of it; the other half is build and test. A wave of AI-powered biofoundries couples generative design with automation to run the design–build–test–learn cycle with minimal hands — one recent setup used a low-cost liquid-handling robot to automate expression, purification and screening of plastic-degrading (PETase) enzymes in 96-well plates. The trajectory points squarely at autonomous protein-engineering platforms. Why it matters for the lab: that loop is REEF’s loop — propose, run, measure, repeat — pointed at molecules instead of cells. The lab that owns the closed loop owns the tempo of discovery.

🛡️ But safeguards are the rare exception

Now the sobering counterweight. Epoch AI cataloged 1,196 biological AI models across nine categories — and found that only 3.2% carry any documented safeguards. (Frontier LLMs are the exception at 95%; for everything else it’s 1.4%.) The census also lands a familiar point: for biology AI, “compute does not appear to be the primary bottleneck” — progress is “more constrained by data availability and quality.” Most models are open, few are risk-assessed. Why it matters for the lab: capability is racing ahead of guardrails, which makes a safety-first, human-in-the-loop posture (REEF’s refusals; FAIR, auditable infrastructure) not a constraint but a differentiator — the exception the field will have to make the rule.

Design is solved-ish, build-and-test is automating fast, and the guardrails are the lagging variable. The interesting work now isn’t just making biology programmable — it’s making it programmable responsibly.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.