BOW-GBDT

BOW-GBDT predicts interactions between G protein–coupled receptors (GPCRs) and small-molecule drugs to support computational prediction of GPCR–drug interactions.

Key Features:

• GPCR feature extraction: A bag-of-words (BOW) model weighted by the Silhouette Coefficient extracts low-dimensional protein sequence features from GPCR sequences.
• Drug feature extraction: Discrete wavelet transform (DWT) is applied to molecular fingerprints to derive drug molecule features.
• Feature integration: Extracted GPCR and drug features are combined into joint representations for modeling.
• Data balancing: SMOTE is used to balance the training dataset and address class imbalance.
• Feature refinement: An artificial neural network further processes and refines the integrated features.
• Prediction model: A gradient boosting decision tree (GBDT) is trained on the refined features for interaction prediction.
• Evaluation datasets: Model performance was assessed using the D92M dataset (cross-validation) and the independent Check390 dataset (independent test).
• Performance claim: The approach demonstrated improved prediction accuracy relative to existing machine learning models on the tested datasets.

Scientific Applications:

• GPCR–drug interaction prediction: Computationally predict binary interactions between GPCRs and small-molecule drugs for hypothesis generation.
• Candidate prioritization for experimental validation: Rank and prioritize GPCR-targeting small molecules for downstream experimental testing in drug discovery.
• Benchmarking machine learning methods: Provide a comparative framework using D92M and Check390 for evaluating ML approaches to GPCR–drug interaction prediction.

Methodology:

Protein sequences are encoded by a bag-of-words model weighted with the Silhouette Coefficient; drug molecular fingerprints are transformed by discrete wavelet transform (DWT); features are combined and balanced with SMOTE; an artificial neural network refines integrated features; a gradient boosting decision tree (GBDT) is trained and evaluated by cross-validation on D92M and independent testing on Check390.