Lab Newsletter — September 2, 2026: What the Fold Won't Tell You

We have gotten very good at guessing what a protein looks like. The headline result of the structure era isn’t one crystal — it’s scale: an ESM protein language model (Lin … Rives, Science, 2023) can “demonstrate direct inference of full atomic-level protein structure from primary sequence,” and its authors used that to build “the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence.” Six hundred million folds — “a view into the vast breadth and diversity of natural proteins.” And yet, staring at that atlas, a biologist still has to ask the question a shape can’t answer: what does this one do? A fold is a hint, not a verdict. Most of those metagenomic proteins are functionally dark. So today’s digest is about the other machine-learning problem — the less glamorous, arguably more useful one — of reading a protein’s job from its letters or its shape.

🔡 From sequence alone

The oldest way to guess a protein’s function is by resemblance: line it up against proteins we already understand and copy their labels. It works until the query has no close, well-annotated relative — which, for the dark proteome, is most of the time. Deep learning offers a different bet: learn the sequence-to-function mapping directly. DeepGOPlus (Kulmanov & Hoehndorf, Bioinformatics, 2020) is the clean version of that idea — a method “for predicting protein functions from sequence alone which combines deep convolutional neural network (CNN) model with sequence similarity based predictions.” It predicts Gene Ontology terms (the standard vocabulary of what a protein does, where, and in what process) and, tellingly, it’s practical: it can “annotate around 40 protein sequences per second on common hardware.” At six hundred million proteins, throughput is not a footnote — it’s the difference between a method you can run on a metagenome and a method you can only demo.

ProteInfer (Sanderson … Colwell, eLife, 2023) pushes the same instinct further and lands on something quietly radical about access. Its networks “directly predict a variety of protein functions — Enzyme Commission (EC) numbers and Gene Ontology (GO) terms — directly from an unaligned amino acid sequence,” no alignment step at all. The authors are careful to frame it as a complement, not a conqueror — the approach “provides precise predictions which complement alignment-based methods” — but the payoff is what a single lightweight network makes possible: “an in-browser graphical interface for protein function prediction in which all computation is performed on the user’s personal computer with no data uploaded to remote servers.” Function prediction that runs in a browser tab, privately, is exactly the kind of open, low-friction tool this lab likes to point at.

🧪 Enzymes: getting the EC number right

Enzymes are the sharpest test of function prediction, because their labels — Enzyme Commission numbers — are precise, hierarchical, and checkable in a test tube. The workhorse for decades has been BLAST: find the closest known enzyme, borrow its EC number. CLEAN (Yu … Zhao, Science, 2023) shows how much room that left on the table. It is “a machine learning algorithm named CLEAN (contrastive learning-enabled enzyme annotation) to assign EC numbers to enzymes with better accuracy, reliability, and sensitivity compared with the state-of-the-art tool BLASTp.” The contrastive-learning trick — teaching the model which enzymes belong together and which don’t — does more than beat a baseline. It “empowers CLEAN to confidently (i) annotate understudied enzymes, (ii) correct mislabeled enzymes, and (iii) identify promiscuous enzymes with two or more EC numbers,” and the authors back those claims “by systematic in silico and in vitro experiments.” That middle point is the one to sit with: a model that doesn’t just fill blanks but finds errors in the existing annotations is doing something a lookup table never could.

🧬 From structure: where the function lives

The structure atlas isn’t a distraction from this problem — it’s a second input to it. Two proteins with unrelated sequences can fold into the same active-site geometry and do the same job; sequence alone would miss the kinship, but structure sees it. DeepFRI (Gligorijević … Bonneau, Nature Communications, 2021) is built for exactly that bridge: “a Graph Convolutional Network for predicting protein functions by leveraging sequence features extracted from a protein language model and protein structures.” Treating the folded protein as a graph, it “outperforms current leading methods and sequence-based Convolutional Neural Networks and scales to the size of current sequence repositories.” And it answers a question the sequence models can’t: where. Through class activation mapping, DeepFRI enables “function predictions at an unprecedented resolution, allowing site-specific annotations at the residue-level in an automated manner” — pointing not just at that a protein binds something, but at which residues do the binding. Fold the metagenome with ESMFold, read function off those folds with DeepFRI, and the two halves of “annotate everything” start to close.

🧭 The honest frontier — and why it’s our kind of problem

Every method above reports a number; the question is who scored it. Protein function prediction has an unusually mature answer: CAFA (Zhou … Friedberg, Genome Biology, 2019), “an ongoing, global, community-driven effort to evaluate and improve the computational annotation of protein function.” The third round did something benchmarks rarely dare: it reached out of silico and into the lab, where “computational predictions and assessment goals drove some of the experimental assays, resulting in new functional annotations for more than 1000 genes.” Predictions that generated real wet-lab experiments — that’s the prove-it standard at its strongest. And CAFA stays honest about the ceiling, concluding that “while predictions of the molecular function and biological process annotations have slightly improved over time, those of the cellular component have not.” Some of this problem is genuinely getting solved; some of it isn’t yet — and a good benchmark says which is which. It’s the same prove-it discipline we flagged in cryo-EM’s CryoBench and yesterday’s tissue-segmentation benchmark: the number that counts is the one measured against a public test you didn’t design.

And it lands where we work. A serious virtual cell or Human Cell Simulator needs a parts list — not just the shape of every protein but its function: what catalyzes what, what binds what, what sits in which compartment. That list is what function prediction writes. It’s the functional partner to the spatial proteome mapped by the Human Protein Atlas next door at KTH — where proteins are, meets what they do. And the way these tools should travel is the ethos we keep returning to: open, callable, benchmarked. ProteInfer’s in-browser, no-upload interface and CLEAN and DeepFRI as callable models are the same spirit as serving models through the BioImage Model Zoo and BioEngine — publish the tool, and publish the test that could embarrass it. We can fold the whole metagenome now. Teaching the machine to say what each of those six hundred million proteins is for is the half that turns an atlas of shapes into a cell you can compute 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.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.