amelie-schreiber/sampling_protein_language_models

Using variations on Gibbs sampling for directed evolution

Variations of Gibbs Sampling of Protein Language Models for Directed Evolution in Silico

This repo uses variations of Gibbs sampling of masked (protein) language models, in particular, ESM-2, in order to sample protein sequence space to create novel proteins. It can be used in conjunction with various scoring methods such as the masked marginal scoring function to form a directed evolution graph with nodes representing proteins, and directed edges indicating fitness of the mutation as scored by the protein language model. This can be used to get a sense of the fitness landscape of mutated proteins.

For example, one can generate multiple mutated proteins from the Gibbs sampling functions, then, using the masked marginal scoring one can determine if the mutations improve or degrade the fitness of the protein to obtain directed edges. One can then sample using the generated proteins to expand the directed graph further.