RAG-ESM: Improving pretrained protein language models via sequence retrieval

Protein language models are significantly advancing the modeling of sequence-function relationships. However, most of them are not directly informed of homology and evolutionary relationships between protein sequences. Here, we propose a method to make them homology-aware. We introduce RAG-ESM, a retrieval-augmented framework that allows to condition pretrained ESM2 protein language models on homologous sequences, using a minimal number of additional cross-attention parameters and minimal computational cost. We show that RAG-ESM models outperform larger ESM2 models for masked amino acid prediction. We find that sequence alignment capabilities spontaneously emerge in specific cross-attention heads of RAG-ESM. By using a discrete diffusion objective for training, and by conditioning on homologs during inference, RAG-ESM reaches state-of-the-art performance for conditional protein sequence generation and motif scaffolding, among sequence-based models. Our method thus possesses strong potential for scalable, efficient and controlled protein engineering.