Graph neural networks (GNNs) are effective at predicting compound properties and activities, but they have two limitations: molecular representation and interpretability. Atom-based molecular graphs are commonly used to represent molecules, but they may not capture important substructures/functional groups that highly impact molecular properties. Interpretability is also important as it can provide scientific insight for optimization, but GNNs are complex and less interpretable. This research introduces techniques to create alternative reduced molecular graph representations that integrate higher-level information and support interpretation. Experiments are conducted on pharmaceutical endpoint datasets, and attention mechanism is used to identify significant substructures. Results show that combining multiple graph representations gives promising performance and provides interpretation aligning with background knowledge. This research could facilitate model understanding and applications in drug discovery.
各種検索
サポート
ヘルプ