MULAN: multimodal protein language model for sequence and structure encoding

Abstract

Motivation: Most protein language models (PLMs) produce high-quality representations using only protein sequences. However, incorporating known protein structures is important for many prediction tasks, leading to increased interest in structure-aware PLMs. Currently, structure-aware PLMs are either trained from scratch or add significant parameter overhead for the structure encoder. Results: In this study, we propose MULAN, a MULtimodal PLM for both sequence and ANgle-based structure encoding. MULAN has a pre-trained sequence encoder and an introduced parameter-efficient Structure Adapter, which are then fused and trained together. Based on the evaluation of nine downstream tasks, MULAN models of various sizes show a quality improvement compared to both sequence-only ESM2 and structure-aware SaProt. The main improvements are shown for the protein–protein interaction prediction (up to 0.12 in AUROC). Importantly, unlike other models, MULAN offers a cheap increase in structural awareness of protein representations because of the finetuning of existing PLMs instead of training from scratch. We perform a detailed analysis of the proposed model and demonstrate its awareness of the protein structure. Availability and implementation: The implementation, training data, and model checkpoints are available at https://github.com/DFrolova/MULAN.