Integrating Functional Knowledge into Protein Design: A Novel Approach to Tokenization and Noise Injection for Function-Aware Protein Language Models

Abstract

Designing novel proteins with specific biological functions remains a fundamental challenge in computational biology. While recent advances in protein language models have enabled powerful sequence-based representations, most models, including state-of-the-art systems like ESM3, fall short in effectively encoding functional context during protein generation. In this work, we present a multimodal protein co-design framework that conditions sequence generation on fine-grained functional annotations, specifically leveraging residue-level Gene Ontology (GO) term labels on sequences from the UniRef100 database. By explicitly associating functional signals with residue elements of proteins, our model learns to generate function-conditioned protein sequences that are biologically plausible and semantically consistent. Unlike prior approaches, which treat function as a secondary feature or a classification task, our method focuses on joint reasoning over function and sequence during the design process. This closes a critical gap in the current landscape of protein design tools, offering a scalable and generalizable approach to co-designing protein sequences with user-specified functional profiles.