E-DBPP: Enhanced DNA-binding protein prediction using an MLP ensemble with optimized embeddings

Abstract

DNA-binding proteins (DBPs) are crucial for various stages of genetic processes, including DNA recombination, repair, and modification. They are widely used in drug discovery because they are crucial parts of antibiotics, steroids, and anticancer drugs. In bioinformatic research, predicting them is the most challenging task. Thus, there is a significant need to create a reliable method for identifying DBPs from sequencing data. This study presents E-DBPP, a Deep Learning (DL) framework for effectively recognizing DBPs from sequence data using an ensemble of four Multi-Layer Perceptron (MLP) classifiers. In the proposed framework, two powerful feature encoding schemes are employed, including ESM2 and FastText, to encode the protein sequence. To achieve this, these features were concatenated and standardized using the StandardScaler. To confirm the effectiveness of the model, 10-fold cross-validation is utilized and compared with other commonly used Machine Learning (ML) and DL models for DBP identification. Furthermore, Local Interpretable Model-agnostic Explanations (LIME) were used for model interpretability to explain individuals’ predictions. On an independent test set, E-DBPP achieved an Area Under the Curve (AUC) of 0.9634, an accuracy of 89.78%, and a Matthews Correlation Coefficient (MCC) of 0.7974, indicating strong and reliable classification performance compared to other state-of-the-art models. These results show that E-DBPP has a reliable predictive ability to classify DBPs.