models/ESM-1v
ESM-1v
ESM-1v is a computer model that has read millions of protein sequences and can guess, without any experiment, whether swapping one building block in a protein will harm how it works.
What it is
Proteins are long chains built from twenty kinds of small unit called amino acids. The order of those units is written down as a sequence, rather like a sentence in an alphabet of twenty letters. A protein language model is trained on huge collections of such sequences, with letters hidden at random, and learns to predict what belongs in each gap. Nothing about chemistry is taught to it directly. It simply picks up which letters tend to sit in which positions, and which combinations never occur in living things.
ESM-1v puts that habit to work on single changes, called mutations. For a given position in a chain, it compares how likely the original letter is against how likely the substitute is. A substitute the model finds surprising is scored as probably damaging; one it finds ordinary is scored as probably tolerated. The score comes from the sequence alone, with no measurement of the protein and no training on experimental results.
Why scientists use it
Testing mutations in a laboratory is slow and costly, and a modest protein already allows thousands of single changes. Scientists use models like this to rank the possibilities first, so that bench work goes to the changes most likely to matter. The same ranking helps in medicine, where a patient may carry a variant of unknown effect, and in protein engineering, where researchers want a chain that is more stable or faster at its job.
Because the model learns general patterns of protein sequence, its internal description of a chain is useful beyond scoring mutations. That description can be reused to compare one protein with another, or as the starting point for a smaller model trained on a specific question, such as which positions in a chain do the chemical work. It stands in for the evolutionary record: what has survived in living things is treated as a guide to what works.
In the record
Limits
The model only knows which sequences look usual. It has no account of the chemistry or shape involved, so it can call a change tolerated when it ruins a particular function, or damaging when it is merely rare. Proteins with few known relatives are judged poorly, because the patterns the model relies on were never well sampled. Its scores rank changes against each other rather than predicting any measured quantity, and they say nothing about effects that depend on the cell, the temperature or the presence of other molecules. Laboratory tests remain necessary.
Written by AIxSci: the background from general knowledge, the examples from checked records.