FuseGO

Evaluating Embedding Fusion Across Species with Unequal Encoder Capacity for Automated Protein Function Prediction

Abstract

Proteins are the workhorses of life, and determining the functions of an uncharacterized protein is a fundamental bioinformatics problem. The function of a protein is defined by a structured vocabulary called Gene Ontology (GO), but determining function in wet labs is highly resource-intensive. Recently, protein language models show promise for function prediction, but it remains unclear whether combining representations improves performance over strong single-model baselines or justifies added complexity. We present an empirical comparison of single-model and fusion-based approaches for predicting functions using protein language models, formulated as a multi-label classification problem.