Abstract
This work investigates Graph Neural Networks
and molecular descriptors average performance
across a range of realistic QSAR problems.
Using 9 GSK internal QSAR datasets, 9 Graph Neural Network layer architectures and their hyperparameters were evaluated on a default set of descriptors. Following this, 5 descriptor classes were evaluated across datasets using a common baseline model architecture. We show that no architecture performed better than any other, and that hyperparameters such as the learning rate, dropout, and message-passing layers are crucial for performance. We likewise show that the choice of molecular descriptors is impactful across datasets. Finally, the impact of tautomers on descriptors showed consistent deviations for expected descriptors, such as the number of hydrogens and graph
centrality measures.
Our recommendations are to direct modelling efforts towards hyperparameter optimization and feature selection rather than focusing on network architecture.
Supplementary materials
Title
Supplementary Information
Description
Details of descriptors studied, statistical tests, and information and analysis on hyperparameters optimisation runs
Actions
Supplementary weblinks
Title
GNN-QSAR GitHub
Description
The GitHub repository containing code and data mentioned in the study
Actions
View