DeepAlloWeb: a web server for interactive allosteric pockets prediction using protein language model

Abstract

Allostery critically influences protein function, presenting unique opportunities in targeted drug development due to reduced side effects compared to orthosteric drugs. This work builds on our previous method, DeepAllo. DeepAlloWeb is an interactive web server to improve user engagement and practical usability for computational biologists and drug discovery researchers. DeepAllo is an advanced computational approach leveraging a fine-tuned protein language model (ProtBERT-BFD) in multitask learning combined with FPocket-extracted features to predict allosteric pockets accurately. Unlike existing allostery prediction servers, which do not utilize protein language models (PLMs), our web server integrates the fine-tuned PLM to achieve better prediction performance and offers an interactive visualization of residue-level attention mechanisms. The DeepAlloWeb resource enables detailed exploration through visualizations of attention mechanisms; an effort for biological interpretability. A case study shows that residue-level attention highlights known allosteric communication in a representative protein, and an additional whole-dataset analysis provides guidance on how layer/head combinations can be interpreted in practice. The webserver may be accessed via: https://3dpath.ku.edu.tr/DeepAllo/.