Abstract
MOTIVATION: Ubiquitination is a crucial post-translational modification that regulates various biological functions, including protein degradation, signal transduction, and cellular homeostasis. Accurate identification of ubiquitination sites is essential for understanding these mechanisms, yet existing prediction tools often lack generalizability across diverse datasets. To address this limitation, we developed Multimodal Ubiquitination Predictor, a deep learning-based approach capable of predicting ubiquitination sites across general, human-specific, and plant-specific datasets. By integrating diverse protein sequence representations-one-hot encoding, embeddings, and physicochemical properties-within a unified deep-learning framework, the proposed method significantly enhances prediction accuracy and robustness, offering a valuable resource for both research and applications in ubiquitination site discovery. RESULTS: Multimodal Ubiquitination Predictor achieved superior performance across general, human-specific, and plant-specific datasets, with 77.25% accuracy, 74.98% sensitivity, 80.67% specificity, an MCC of 0.54, and an AUC of 0.87 on an independent human ubiquitination test dataset. It outperformed existing methods, demonstrating enhanced reliability for ubiquitination site prediction. This robust predictor and dataset serve as valuable resources for future research and discovery. AVAILABILITY AND IMPLEMENTATION: The developed tool, programs, training, and test dataset are available at https://github.com/PakhrinLab/MMUbiPred.