Lab Newsletter — June 28, 2026: Designing Proteins, Reading Genomes

A change of scenery from this week’s agents-and-microscopes run: today the spotlight is on designing and reading biology’s sequences — proteins and genomes.

🧬 Protein design grows into an engineering discipline

A fresh Frontiers editorial (June 25) and a Communications Biology commentary both make the same point: AI protein design has moved from “remarkable demo” to scalable engineering. Generative tools like RFdiffusion and BindCraft now produce high-affinity binders, enzymes and therapeutic candidates with high experimental success rates — and the next frontiers are sharpening: predicting full conformational ensembles (approximating Boltzmann-weighted states at a fraction of molecular-dynamics cost) and making de novo binder design a routine, reliable pipeline. Why it matters for the lab: protein design is the “design” half of the AI-for-biology loop that complements our imaging and cell-modeling work — and the framing we keep coming back to (turn a craft into a reliable, benchmarked engineering discipline) is exactly our north star for agentic tooling, too.

🧪 Evo 2 reads genomes at scale — but scale isn’t everything

Arc Institute’s Evo 2 (Nature, March 2026) is a genomic foundation model at a striking scale: 40 billion parameters, a 1-megabase context, trained on 9 trillion+ nucleotides across eukaryotic and prokaryotic genomes, for generalist prediction and design across DNA, RNA and protein at single-nucleotide resolution. Yet a benchmark in Nature Communications keeps it honest: general-purpose DNA foundation models were competitive at pathogenic-variant identification but lagged specialized models at predicting gene expression and pinning down causal QTLs. Why it matters for the lab: it’s the same lesson the virtual-cell field is learning — scale is necessary but not sufficient; the win comes from matching the model (and a fair benchmark) to the question.

🛡️ The dual-use shadow worth naming

A Frontiers analysis (March 2026) raises a concern the field can’t wave away: AI can now generate proteins that are functionally equivalent to known toxins while sharing almost no sequence similarity — which makes today’s homology-based biosecurity screening effectively blind to them. Why it matters for the lab: as we build more autonomous, agentic tools, responsibility scales with capability. Provenance, guardrails and human-in-the-loop checks aren’t friction — they’re part of doing powerful science well, and a thread we’ll keep pulling on.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. Have lab news to share — a talk, paper, conference or release? Message me on Slack.