Deep-WET: a deep learning-based approach for predicting DNA-binding proteins using word embedding techniques with...

Area of Science:

• Computational biology
• Bioinformatics
• Proteomics

Background:

• DNA-binding proteins (DBPs) are crucial for genetic processes like DNA repair and modification.
• DBPs are important targets in drug discovery for antibiotics and anticancer agents.
• Experimental DBP identification is costly and can be biased, necessitating computational solutions.

Purpose of the Study:

• To develop a novel, accurate, and rapid computational method for identifying DNA-binding proteins (DBPs) from primary sequence information.
• To leverage deep learning and advanced feature engineering for enhanced DBP prediction.

Main Methods:

• Proposed Deep-WET, a deep learning model utilizing Global Vectors, Word2Vec, and fastText for protein sequence encoding.
• Employed differential evolution (DE) for feature weighting and SHapley Additive exPlanations (SHAP) for feature selection.
• Integrated an optimal feature subset into convolutional neural networks (CNNs) for the final predictor.

Main Results:

• Deep-WET demonstrated superior predictive performance over conventional machine learning classifiers in cross-validation and independent tests.
• Achieved high accuracy (78.08%), MCC (0.559), and AUC (0.805) in extensive independent testing.
• Outperformed several state-of-the-art methods for DNA-binding protein prediction.

Conclusions:

• Deep-WET exhibits significant predictive capacity for identifying DNA-binding proteins.
• The developed method provides a valuable tool for large-scale DBP identification in proteomics research.
• A web server and datasets are publicly available to support the scientific community.