Lab Newsletter — August 18, 2026: The Proteins Evolution Never Wrote

For a decade the marquee problem in computational biology ran one direction: given a protein’s sequence, predict its shape. AlphaFold nailed it. But the dream underneath was always the other direction — start from what you want and write the protein that does it. That inverse problem, de novo protein design, is where generative AI has quietly become spectacular: you specify a fold, a binding target, an active site, sometimes just a sentence, and a model invents a protein — often one evolution never got around to making. It’s the protein analysis of the past two weeks turned inside out: not reading molecules, but authoring them.

🧬 The inverse problem, learned

Three tools define the generative toolkit. ProteinMPNN (Dauparas et al., Baker Lab; Science, 2022) solves inverse folding — hand it a target backbone, it designs a sequence that folds to it — and it doesn’t just win on paper: 52.4% native-sequence recovery versus 32.9% for the physics-based Rosetta, and, the headline, it rescued designs that had failed with older methods, confirmed by X-ray and cryo-EM. RFdiffusion (Watson et al., Baker Lab; Nature, 2023) brought diffusion to protein backbones — the same denoising idea behind image generators, fine-tuned from RoseTTAFold — and generated protein binders, symmetric oligomers, enzyme active-site scaffolds and metal-binding proteins; hundreds were made in the lab, and a designed anti-influenza binder matched its blueprint by cryo-EM. And Chroma (Ingraham et al., senior author Grigoryan; Nature, 2023) made design programmable: condition generation on symmetry, shape, substructure, semantics — even a natural-language prompt — and sample proteins to order. Its team experimentally characterized 310 designed proteins, and found them highly expressed and well-folded. Design stopped being a boutique craft and became something you can run.

🗣️ Language as the substrate

The next move made design feel eerily like prompting. ESM3 (Hayes et al., EvolutionaryScale; Science, 2025) is a multimodal generative protein language model — it reasons jointly over sequence, structure and function, so you can mask any of them, give it a partial spec, and have it fill in a coherent protein. Scaled to 98 billion parameters, it did something that made the field sit up: it generated esmGFP, a new, bright green fluorescent protein only 58% identical to the nearest known fluorescent protein — a molecule the authors estimate is equivalent to more than 500 million years of natural evolutionary divergence. Not a protein evolution made, but one it plausibly could have — reached in a single sampling run. And crucially, the team released ESM3-open weights and put the esmGFP sequence in the public domain: the same open, FAIR, callable-model ethos the lab builds its platforms on, now for generative structural biology.

⚗️ From backbone to catalysis — and the prove-it loop

Two things mark how far this has come. First, the 2024 Nobel Prize in Chemistry went half to David Baker “for computational protein design” and half to Demis Hassabis and John Jumper “for protein structure prediction” (AlphaFold) — design and prediction crowned side by side, a field’s coming-of-age. Second, the bar has moved from making a shape to making a function: a 2025 Nature paper, “Computational design of metallohydrolases” (Kim et al., Baker Lab), used the RFdiffusion2 method to design de novo zinc metallohydrolase enzymes from quantum-chemistry active-site geometries — with catalytic activity and crystal structures confirmed in the wet lab. The method itself (a bioRxiv preprint) reports scaffolding all 41 of 41 benchmark active sites in silico, against 16 of 41 for the prior best. Making a fold is table stakes now; making a catalyst is the new frontier.

Here’s why it lands for us. A designed sequence, however confident the model, is a hypothesis about matter — it isn’t a protein until it’s synthesized, expressed and assayed. That is exactly the design → build → test loop, and the lab lives on the build-and-test half: the open model serving that makes these large design models callable, and the autonomous experimentation and agentic-science infrastructure that turns a generated sequence into a plate, a reading, a result. Generative design is one of the great AI-for-science stories — proteins evolution never wrote, drawn on demand — and it carries the same prove-it discipline the digest keeps returning to: the design is the easy, beautiful part; the world still has to agree.

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; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.