Drug-Target Interaction Prediction: A Comparative Study of Molecular Fingerprints and Graph Neural Networks

This study focuses on activity prediction of Epidermal Growth Factor Receptor (EGFR) inhibitors. Based on data from the ChEMBL database, it compares two methods: Morgan molecular fingerprints + Random Forest and Graph Neural Networks (GNNs). A complete machine learning workflow is implemented using tools like RDKit, PyTorch Geometric, and SHAP. Results show that traditional methods perform better under the current data scale.

Drug-Target Interaction (DTI) prediction is a core issue in drug discovery, which can accelerate new drug development and reduce experimental costs. This project targets EGFR inhibitor activity prediction; as a key target for cancer treatment, the development of EGFR inhibitors is of great significance for tumor therapy.

Data is sourced from the ChEMBL database (CHEMBL203), with 17,723 original records. After cleaning, 8,728 compounds are obtained (89% active, 11% inactive). Data features include an average pIC50 of 7.23 and an average molecular weight of 488.1 Da. The cleaning process: ChEMBL API acquisition → IC50 extraction → pIC50 conversion → binary label generation.

Method 1: Morgan Fingerprints + Random Forest

• Molecular representation: Morgan fingerprints generated by RDKit (radius 2, 2048-bit binary vector)
• Model: Random Forest (200 trees, class_weight='balanced', 5-fold cross-validation)
• Interpretability: SHAP analysis for feature importance

Method 2: Graph Neural Network

• Molecular representation: Atoms as nodes (15-dimensional features), chemical bonds as edges
• Model: 3-layer GCN, global average pooling, Adam optimizer, trained for 50 epochs

Results show that the Random Forest has an ROC-AUC of 0.9694, while the GNN has 0.8887. Reasons include: 1. Class imbalance: RF uses class_weight to mitigate it, but GNN does not; 2. Data scale: 8,728 samples are small for GNN; 3. GNN training did not converge (AUC still increased after 50 epochs); 4. Morgan fingerprints are more effective in small to medium datasets.

Tools: RDKit (molecular processing), scikit-learn (RF implementation), PyTorch/PyTorch Geometric (GNN), ChEMBL API (data acquisition), SHAP (interpretability), etc. Project Structure: Includes directories like data (raw/processed), notebooks (6 phase scripts), src (function functions), models (model files), etc.

1. Method selection should consider data scale: traditional methods are better for small to medium datasets; 2. Class imbalance needs to be addressed: a common issue in biological activity data; 3. Interpretability is important: SHAP analysis guides compound design; 4. Value of complete workflow: conducive to reproducible research.

This project provides a complete workflow for EGFR inhibitor activity prediction. Comparing the two methods, it is found that although GNNs are theoretically suitable for complex structure learning, Morgan fingerprints + Random Forest perform better under the current data. This suggests that researchers should choose methods based on practical problems rather than blindly pursuing new ones.