HRGCN+

HRGCN+ predicts physicochemical and biological properties for QSAR modeling in drug discovery by combining molecular graphs and molecular descriptors within a hyperbolic relational graph convolutional network.

Key Features:

• Integration of Molecular Graphs and Descriptors: Combines molecular graphs with traditional molecular descriptors as inputs to a modified graph neural network.
• Hyperbolic Relational Graph Convolutional Architecture: Implements a hyperbolic relational graph convolutional network (HRGCN) with modifications to process combined graph and descriptor information.
• Performance on Drug-Discovery Datasets: Demonstrated superior performance across 11 drug-discovery-related datasets.
• Enhanced Predictive Power: Incorporating molecular descriptors into graph-based methods increases predictive accuracy and robustness.
• Noise Resistance: Exhibits anti-noise capabilities to maintain reliable performance on complex or imperfect datasets.
• Interpretability: Provides interpretative insights at both the atom and descriptor levels to identify contributions to predicted activities.

Scientific Applications:

• Initial Screening: Predicts compound properties to support initial screening and identification of promising candidates.
• Lead Optimization: Supports lead optimization by predicting bioactivities and physicochemical properties during compound modification.
• Mechanistic Interpretation: Uses atom- and descriptor-level interpretability to inform rational design and modification of compounds.

Methodology:

A modified hyperbolic relational graph convolutional network processes molecular graphs enriched with molecular descriptor data.