Fluorescent Protein Made by a Protein Language Model Resembles Nature's Only 58%

A protein sequence can be seen as a kind of sentence left by billions of years of evolution. Researchers at EvolutionaryScale and the Arc Institute in the United States reported in Science ESM3, a generative language model trained on these sentences, and said it actually produced functional proteins that do not exist in nature.

Sequence, structure and function in one model

ESM3 is a multimodal generative model that handles not only amino acid sequences but also three-dimensional structure and functional annotations within the same model. When a user gives mixed instructions, such as 'keep this structure while having this function,' it generates a matching sequence, and additional alignment training can improve how well it follows instructions.

The researchers instructed it to make a fluorescent protein, then synthesized and tested the candidates it generated. They obtained a bright fluorescent protein with only 58% sequence identity to known fluorescent proteins, that is, one that is phylogenetically far removed. The authors estimated that natural evolution would take about 500 million years to open up that much distance.

Limitations

The '500 million years' figure is an estimate converted from distances between natural sequences, and the successful cases are some of the candidates that were synthesized and tested. The result comes from fluorescence, a function that is relatively easy to measure, so generalizing it to enzyme activity or therapeutic proteins requires more validation.