DisPredict3.0: Prediction of intrinsically disordered regions/proteins using protein language model

Abstract

Intrinsically disordered proteins (IDPs) or protein regions (IDRs) do not have a stable three-dimensional structure, even though they exhibit important biological functions. They are structurally and functionally very different from ordered proteins and can cause many critical diseases. Accurate identification of disordered proteins/regions significantly impacts fields such as drug design, protein engineering, protein design, and related research. However, experimental identification of IDRs is complex and time-consuming, necessitating the development of an accurate and efficient computational method. The recent development of deep learning methods for protein language models shows the ability to learn evolutionary information from billions of protein sequences. This motivates us to develop a computational method, named DisPredict3.0, to predict proteins’ disordered regions (IDRs) using evolutionary information from a protein language model. Compared to the state-of-the-art method in the CAID (2018) assessment, DisPredict3.0 has an improvement of 2.51 %, 16.13 %, 17.98 %, and 11.94 % in terms of AUC, F1-score, MCC, and kappa, respectively. In addition, in the CAID-2 assessment (2022), DisPredict3.0 shows promising results and is ranked first for disorder residue prediction on the Disorder-NOX dataset. The DisPredict3.0 webserver is available at https://bmll.cs.uno.edu.