ESMFold · AIxSci

What it is

Proteins are long chains of small units called amino acids, and each chain folds into a particular shape. The shape is what lets a protein do its job, so biologists want to know it. ESMFold predicts that shape. It is built on top of a protein language model called ESM-2, a program trained on huge numbers of protein sequences in much the way a text model is trained on sentences, learning which amino acids tend to sit together. Those learned patterns carry information about folding. A second part of the program turns them into coordinates: a guess at where every atom sits in space.

What sets ESMFold apart from some other predictors is what it leaves out. Many methods begin by searching databases for related sequences from other species and lining them up, because positions that change together often touch in the folded shape. That search is slow. ESMFold skips it and reads one sequence on its own, relying on what the language model already absorbed during training. Because it is fast, it has been run across very large collections of sequences in advance, and the resulting shapes published in a public set called the ESM Metagenomic Atlas, which others can simply download.

Why scientists use it

Measuring a protein's shape in the laboratory is slow and expensive, and many proteins resist the attempt altogether. Prediction stands in for that measurement. It gives researchers a shape to work with when none has been measured, which matters most for proteins from poorly studied organisms, where few or no relatives are known and alignment-based methods have little to go on.

Speed also changes what questions can be asked. A predictor cheap enough to run on millions of sequences turns structure from something you look up for one protein into something available for a whole collection. Predicted shapes then feed other tools: programs that compare shapes to find distant cousins, programs that judge whether a mutation will destabilise a protein, and programs that guess what a protein does.

In the record

structural-biologyAI-labelled protein pairs used to test how well function predictors handle unknown proteinsPre-computed ESMFold structures from the ESM Metagenomic Atlas supplied the shapes for a filtered set of orphan proteins, which were then aligned against each other.structural-biologyNeural networks trained on atomic surroundings score how mutations change proteinsStructures were obtained through the ESM Metagenomic Atlas interface, then pre-processed to add hydrogens, charges and surface areas before each residue's surroundings were encoded.structural-biologyA shared test set for predicting which enzyme carries out a reactionRecorded as a structure source alongside enzyme sequences annotated with reaction classes, which were clustered into held-out training and testing splits.

Limits

A prediction is a guess, not a measurement. Skipping the search for related sequences makes ESMFold fast but usually less accurate than methods that use it, especially for proteins with unusual folds. It returns one fixed shape, while real proteins flex, shift and respond to partners, and it says nothing about what a protein does. Regions with no stable shape at all may still be drawn as though they had one. Confidence scores help, but a wrong structure that looks plausible can quietly mislead everything built on top of it.

Written by AIxSci: the background from general knowledge, the examples from checked records.