Hybrid ensembles and transformer-based models for predicting protein acetylation sites

Predicting post-translation modification (PTM) sites, particularly acetylation sites, is an important research area in bio-informatics and computational biology. Traditional machine learning algorithms, while effective, often struggle with over-fitting and generalization. This research aims to develop a hybrid ensemble and transformer models for acetylation site prediction to address these challenges. The hybrid ensemble model combines the strengths of XGBoost, Random Forest, Rotation Forest, and Convolutional Neural Network (CNN) models to improve prediction accuracy. The transformer model is based on the Transformer architecture and is designed to capture long-range dependencies in protein sequences. We evaluate the performance of both models using cross-validation on our dataset. The hybrid ensemble model outperforms individual models and existing methods, achieving reasonable performance. However, the transformer model’s performance is significantly lower than the hybrid ensemble model, indicating that it may not be well-suited for acetylation site prediction. Future research directions include dataset augmentation, transformer model improvements, feature engineering, and interpretable models to improve prediction accuracy further.