PROTGOAT : Improved automated protein function predictions using Protein Language Models

Abstract Accurate prediction of protein function is crucial for understanding biological processes and various disease mechanisms. Current methods for protein function prediction relies primarily on sequence similarities and often misses out on important aspects of protein function. New developments in protein function prediction methods have recently shown exciting progress via the use of large transformer-based Protein Language Models (PLMs) that allow for the capture of nuanced relationships between amino acids in protein sequences which are crucial for understanding their function. This has enabled an unprecedented level of accuracy in predicting the functions of previously little understood proteins. We here developed an ensemble method called PROTGOAT based on embeddings extracted from multiple and diverse pre-trained PLMs and existing text information about the protein in published literature. PROTGOAT outperforms most current state-of-the-art methods, ranking fourth in the Critical Assessment of Functional Annotation (CAFA 5), a global competition benchmarking such developments among 1600 methods tested. The high performance of our method demonstrates how protein function prediction can be improved through the use of an ensemble of diverse PLMs. PROTGOAT is publicly available for academic use and can be accessed here: https://github.com/zongmingchua/cafa5