ISUMsite: Identification of sumoylation sites based on Smote-ENN resampling technique

Sumoylation is a reversible post-translational modification that plays a crucial role in regulating various biochemical functions of proteins. Dysregulation of sumoylation has been implicated in several diseases, including cancer and neurodegenerative disorders. Identification of sumoylation sites can offer insights into the underlying mechanisms of these diseases and potentially uncover new targets and strategies for their diagnosis and treatment. Nevertheless, current research faces several challenges. Primarily, the training datasets suffer from imbalance, causing models to exhibit bias towards predicting majority categories while performing poorly on minority categories. Additionally, reliance on a single feature representation or a single classifier may limit performance and robustness when dealing with heterogeneous datasets. This study introduces an updated ISUMsite framework for identifying sumoylation sites by integrating handcrafted residue-level descriptors with pre-trained protein language model representations, and further improving model performance via data-driven optimization. For Dataset 1, we retained the distance-based residue feature extraction strategy (DR) and introduced ProteinGLM embeddings as complementary representations. A genetic algorithm (GA) was then employed to jointly optimize classifier hyperparameters and fusion weights for weighted probability fusion. The optimal fusion weights were DR:0.025 and ProteinGLM:0.975, achieving Acc, Sp, Sn, MCC, and AUC values of 99.78%, 99.58%, 99.98%, 0.9956, and for Dataset 2, because DR performed poorly and the TDA-based pathway was removed due to methodological limitations, we trained a ProteinGLM-only predictor and used RandomizedSearchCV (RSCV) with accuracy-first scoring to search for optimal hyperparameters. The final model achieved Acc, Sp, Sn, MCC, and AUC values of 87.65%, 94.29%, 81.01%, 0.7597, and 0.9254 on the independent test set. For ease of use, the prediction tools and datasets developed in this study are available at https://github.com/wmqskr/ISUMsite.