ESM-IF1

What it is

Proteins are chains of small chemical units called amino acids. The order of those units, the sequence, decides the shape the chain folds into, and the shape decides what the protein does. Predicting shape from sequence is called folding. Going the other way, from a shape you want to a sequence that would produce it, is called inverse folding. ESM-IF1 is a model trained for that second job. It is shown a backbone, the bare framework of a protein's fold, and it suggests which amino acid belongs at each position. Because it was trained on many known structures, it has absorbed statistical regularities about which units tend to sit where in a given local geometry.

Why scientists use it

A single shape can be made by an enormous number of different sequences, far too many to test one by one in a laboratory. A model that reads a shape and scores or proposes sequences narrows that space cheaply. Researchers use it in two broad ways. They use it to design: give the model a target shape, take its suggestions, and build those proteins. They also use it to judge: ask how well a given sequence fits a given structure, and read that fit as a rough measure of how comfortable, and so how stable, the protein is. The same model can therefore serve design work and assessment work.

In the record

Limits

The model learns what is common in known structures, so it leans towards the familiar and may do poorly on unusual folds or on parts of proteins that are flexible rather than fixed. A good fit score is not a measurement of stability or of function; it is a statistical preference, and proteins that score well can still fail in the laboratory. It also depends on the input shape being right, so errors in a predicted structure pass straight through into its suggestions.

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

Model family

• Protein language model

Every article