Deep-WET

Deep-WET predicts DNA-binding proteins (DBPs) from primary protein sequence using deep learning to support identification of proteins involved in recombination, repair, modification and applications in drug discovery for steroids, antibiotics, and anticancer agents.

Key Features:

• Primary sequence encoding: Uses primary protein sequence information as the input basis for analysis.
• Embedding models: Encodes protein sequences using Global Vectors (GloVe), Word2Vec, and fastText to generate sequence representations.
• Differential Evolution weighting: Applies a differential evolution (DE) algorithm to assign weights to encoded features.
• SHAP-based feature selection: Employs SHapley Additive exPlanations (SHAP) to identify and remove irrelevant features, retaining an optimal feature subset.
• Convolutional Neural Networks: Feeds the selected feature subset into convolutional neural networks (CNNs) for prediction of DBPs.
• Performance evaluation: Evaluates model performance via cross-validation and independent tests, reporting accuracy 78.08%, Matthews correlation coefficient (MCC) 0.559, and area under the curve (AUC) 0.805.

Scientific Applications:

• DBP identification: Identification of DNA-binding proteins from sequence data.
• Proteomics studies: Investigation of proteins involved in recombination, repair, and modification.
• Drug discovery support: Support of drug discovery efforts targeting steroids, antibiotics, and anticancer agents through DBP identification.
• Large-scale screening: Large-scale identification and screening of candidate DBPs in proteomics datasets.

Methodology:

Encode primary sequence with Global Vectors (GloVe), Word2Vec, and fastText; weight encoded features using differential evolution (DE); perform feature selection with SHapley Additive exPlanations (SHAP); train convolutional neural networks (CNNs); and evaluate using cross-validation and independent tests.