Graph Neural Networks with Multi-features for Predicting Cocrystals using APIs and Coformers Interactions

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Materials Science

Background:

• Active pharmaceutical ingredients (APIs) are increasingly explored for their solid dosage forms.
• Pharmaceutical cocrystals offer an attractive route for drug substance development, guided by FDA approvals.
• Identifying suitable coformers for API cocrystal formation is a significant challenge.

Purpose of the Study:

• To develop and implement graph neural networks (GNNs) for predicting API-coformer cocrystal formation.
• To compare the performance of GNNs against traditional descriptor-based models.
• To introduce a novel API-coformers relational graph dataset.

Main Methods:

• Implementation of Graph Convolutional Networks (GCN), GraphSAGE, and Relational Graph Convolutional Networks (RGCN).
• Utilized a newly introduced API-coformers relational graph dataset for training and validation.
• Compared GNN performance against Random Forest, Support Vector Machine, Extreme Gradient Boosting, and Artificial Neural Networks.

Main Results:

• All implemented GNN models demonstrated high prediction accuracies: GCN (91.36%), GraphSAGE (94.60%), and RGCN (95.95%).
• RGCN outperformed other models by effectively capturing complex interactions and relationships between APIs and coformers.
• The models adeptly learned the topological structure inherent in the graph data.

Conclusions:

• GNNs provide a powerful and efficient approach for predicting pharmaceutical cocrystal formation.
• RGCN shows particular promise due to its ability to model intricate relationships crucial for cocrystal prediction.
• This computational strategy can accelerate the drug discovery and development process by optimizing coformer screening.