A Protein Structure Enhanced Pre-training Model for Allosteric Site Detection

Protein allostery, one of the crucial mechanisms in which proteins play and regulate their functions, is a biological process that affects active sites by perturbing allosteric sites. The recognition of allosteric sites holds vital importance in drug design and enzyme engineering. Consequently, detecting potential allosteric sites is increasingly drawing widespread attention in the realm of allosteric research. Several approaches have been proposed for predicting allosteric sites based on computer simulation techniques and traditional machine learning algorithms using pocket features. However, these methods either require extremely high computational resources and long sampling time or rely heavily on designed features. In this paper, we propose a novel method consisting of Protein Structure Bidirectional Encoder Representations from Transformers (PsBERT) and Protein Structure Masked Language Model (PsMLM) Pre-training Task to detect allosteric sites. Our model aims to enhance the encoder's comprehension of protein structures through the pre-training task, thus achieving the goal of predicting allosteric sites based on sequence information and coordinate information. Extensive experiments we conducted on the ASD2023 dataset demonstrate the effectiveness of our model.