Improved GNNs for Log D_7.4 Prediction by Transferring Knowledge from Low-Fidelity Data.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Quantitative structure-property relationships (QSPR)

Background:

• Lipophilicity, indicated by the n-octanol/buffer distribution coefficient (log D7.4), is crucial for predicting ADMET properties and druggability.
• Graph neural networks (GNNs) can extract features from molecular graphs to learn structure-property relationships (SPRs) but are limited by small datasets.
• Existing GNN models often struggle with accurate log D7.4 prediction due to data limitations.

Purpose of the Study:

• To develop an enhanced transfer learning strategy (PCFE) to improve GNN performance for log D7.4 prediction.
• To leverage large-scale computational data for pretraining and high-fidelity experimental data for fine-tuning GNN models.
• To create a robust and accurate predictive model for log D7.4 accessible via a webserver.

Main Methods:

• Implemented a transfer learning strategy (PCFE): pretraining GNNs on 1.71 million computational log D data, followed by fine-tuning on 19,155 experimental log D7.4 data.
• Evaluated three GNN architectures (GCN, GAT, Attentive FP) and compared the best model (cx-Attentive FP) against descriptor-based models (RF, GB, SVM, XGBoost).
• Employed Shapley additive explanations (SHAP) for feature importance, attention mechanisms for substructure identification, and matched molecular pair analysis (MMPA) for substituent contributions.

Main Results:

• The PCFE strategy significantly improved GNN performance for log D7.4 prediction across tested architectures.
• The optimal cx-Attentive FP model achieved a high test accuracy (R_test^2 = 0.909), outperforming four established descriptor-based models.
• The study identified key descriptors and substructures influencing log D7.4 and confirmed model robustness across different data sizes and splitting strategies.

Conclusions:

• The cx-Attentive FP model, trained using the PCFE strategy, provides a reliable and accurate tool for predicting log D7.4.
• The developed webserver offers free log D7.4 prediction services, aiding researchers in drug discovery.
• Pretraining GNNs on low-fidelity data shows promise for improving predictions of other critical endpoints in drug discovery.