ProCAS: Multi-Lysine Post-Translational Modification Site Prediction Based on the SPUS Algorithm and the ProCAS Architecture.

Abstract

Among PTMs, multiple chemical modifications on lysine residues have received extensive attention due to their dynamic and reversible properties. Although mass spectrometry can accurately identify modification sites, its high technical requirements and limited throughput restrict large scale application. In recent years, advances in deep learning have provided new strategies for high throughput prediction of lysine modifications. However, most existing methods still face limitations in simultaneous prediction of multiple lysine modification types, protein-level contextual representation, and prediction bias caused by imbalanced data distributions. To address these challenges, we developed an intelligent prediction system termed ProCAS. This system integrates multi-dimensional feature engineering, including: (1) an MLPSTAAP feature matrix derived from sequence information; (2) CTD descriptors that represent physicochemical properties of protein sequences; and (3) deep representations generated by ESM-2. A specially designed SPUS algorithm performs dynamic balancing optimization at the data distribution level, thereby increasing the modelling contribution of rare modification samples. The ProCAS neural architecture combines four core components, including bidirectional LSTM, a Self-Highway gating module, channel attention, and multi-head self-attention mechanisms. This design enables effective capture of both local and global sequence context and their complex associations with modification related functions. A bilinear mapping classifier is further applied to jointly output the probability distributions of four lysine modification types. Detailed ablation analysis demonstrated that the SPUS module plays a decisive role in alleviating data imbalance, while the ProCAS architecture effectively establishes deep associations between sequence features and modification types. In addition, SHAP based interpretability analysis was conducted to systematically examine the decision mechanisms of multi-label prediction. The results showed that ESM-2 features provided clear label-specific contributions across different modification types, while CTD features captured global physicochemical property patterns, together explaining both local sequence signals and broader physicochemical determinants of lysine modifications. Code and datasets are available at https://github.com/Lilyjoys/ProCAS, and a free web platform is provided at http://www.mlysprocas.com.