Lab Newsletter — September 19, 2026: A Structure for Every Sequence

This week we’ve followed proteins in every direction — designing antibodies, locating them in the cell, reading their function. All of it rests on one thing we’ve quietly taken for granted all month: that you can go from a protein’s amino-acid sequence to its 3D shape with a computer. For half a century that was biology’s most famous unsolved problem — the “protein folding problem.” Today’s digest goes back to the breakthrough itself: the handful of models that turned sequence-to-structure from a grand challenge into a lookup, and in doing so laid the foundation almost every other digest this month stands on.

🧬 The breakthrough: atomic accuracy, at last

The turning point came at the CASP14 blind assessment in 2020, and landed in print the next year. Jumper et al. (Nature, 2021) opened AlphaFold with the plainest possible stakes: “proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function.” Their result — “highly accurate protein structure prediction with AlphaFold” — reached, for the first time, accuracy competitive with experiment across a huge range of targets, from sequence alone. Decades of X-ray crystallography, NMR, and cryo-EM had mapped structures one hard-won protein at a time; suddenly a model could predict them in hours. It is hard to overstate the shift: the field’s defining problem became, for most proteins, effectively solved.

🔗 The open parallel: three tracks, and complexes for free

A breakthrough only becomes a movement when the community can build on it. Days after CASP14, Baek et al. (Science, 2021) delivered RoseTTAFold, an open method that “explored network architectures that incorporate related ideas” and found “the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated.” It reached “accuracies approaching those of DeepMind in CASP14” — and threw in a bonus that mattered enormously: “rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking.” Crucially, the authors “make the method available to the scientific community to speed biological research.” Open weights, open science — the pattern the lab bets on.

🌍 The scale jump: folding a whole proteome

With a fast, accurate predictor, the natural next move is everything at once. Tunyasuvunakool et al. (Nature, 2021) turned AlphaFold on an entire species. Their framing captures how sparse structural knowledge really was: “after decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally determined structure.” Their answer — “highly accurate protein structure prediction for the human proteome” — filled in the other 83%, delivering confident structural models across nearly the whole set of human proteins. Overnight, “do we have a structure for this protein?” stopped being a research project and became a query.

📚 Make it open: a database for all of protein space

Predictions only change a field when everyone can reach them. Varadi et al. (Nucleic Acids Research, 2022) built the AlphaFold Protein Structure Database — “an openly accessible, extensive database of high-accuracy protein-structure predictions.” It ships not just coordinates but calibrated confidence: “per-residue and pairwise model-confidence estimates and predicted aligned errors.” And the scale is staggering — the initial release held “over 360,000 predicted structures across 21 model-organism proteomes,” slated to expand “to cover most of the (over 100 million) representative sequences from the UniRef90 data set.” A structural map of almost all known proteins, free to anyone — the same open-infrastructure spirit as the lab’s BioImage Model Zoo and BioEngine, applied to structure.

🔤 The language-model route: no alignment required

AlphaFold and RoseTTAFold both lean on multiple-sequence alignments — powerful, but slow to build and thin for “orphan” proteins. Lin et al. (Science, 2023) showed another way with ESMFold: “direct inference of full atomic-level protein structure from primary sequence using a large language model.” Their observation is remarkable — “as language models of protein sequences are scaled up to 15 billion parameters, an atomic-resolution picture of protein structure emerges in the learned representations.” The payoff is speed: “an order-of-magnitude acceleration of high-resolution structure prediction,” fast enough to fold the unmapped microbial world. They used it to build “the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence.” Structure, straight from the language of life.

🧩 Beyond the monomer: predicting interactions

A cell is not a bag of lone proteins — it runs on complexes: proteins with DNA, RNA, ligands, ions. Abramson et al. (Nature, 2024) generalized the whole approach with AlphaFold 3 — “accurate structure prediction of biomolecular interactions with AlphaFold 3.” Rather than a protein-only folder, it is a single model that predicts the joint 3D structure of many molecule types together, bringing drug-like ligands and nucleic acids into the same unified framework. The question shifts from “what shape is this protein?” to “what does this molecular machine look like when its parts come together?” — the shape of function itself.

🧫 Why it’s our kind of problem

Read across the six and it’s the bedrock of nearly everything the lab digests. Structure prediction is now a primitive — fast, accurate, and open — which is exactly why the frontier has moved to what you do with a structure: predict function, design new proteins and antibodies, model motion and complexes. The two design choices that made it stick are the lab’s own creed: open weights and open databases (RoseTTAFold, ESMFold, AlphaFold DB) turn a result into shared infrastructure, and learned representations (the ESM language model growing structure “for free” as it scales) are the same bet the lab makes across imaging and single-cell foundation models. Above all, structure is a load-bearing layer for the virtual cell: you cannot simulate a molecular machine you cannot see, and now — for almost every protein, and increasingly for their partners — we can. A structure for every sequence is not the end of the story; it’s the substrate the rest of the story is written on.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.