Hyperbolic relational graph convolution networks plus: a simple but highly efficient QSAR-modeling method.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Quantitative structure-activity relationship (QSAR) methods are crucial for drug discovery.
• Existing QSAR methods are primarily descriptor-based or graph-based.
• There is a need for improved predictive models in drug discovery.

Purpose of the Study:

• To develop a highly efficient modeling method for predicting compound druggability and bioactivities.
• To combine molecular graphs and descriptors for enhanced QSAR modeling.
• To improve the accuracy and interpretability of drug discovery predictions.

Main Methods:

• Developed a modified graph neural network, hyperbolic relational graph convolution network plus (HRGCN+).
• Integrated molecular graphs and traditional molecular descriptors as input features.
• Evaluated HRGCN+ performance on 11 drug-discovery-related datasets.

Main Results:

• HRGCN+ achieved state-of-the-art performance across 11 datasets.
• The addition of molecular descriptors enhanced the predictive power of graph-based methods.
• The method demonstrated strong anti-noise capabilities and provided atom- and descriptor-level interpretability.

Conclusions:

• HRGCN+ offers a powerful and efficient approach for drug discovery predictions.
• Combining graph and descriptor information boosts predictive accuracy.
• The method aids medicinal chemists in extracting insights from complex data.