Comparison of Atom Representations in Graph Neural Networks for Molecular Property Prediction

TL;DR

This paper investigates how different atom representations affect the performance of graph neural networks in predicting molecular properties, highlighting the importance of featurization in chemical analysis.

Contribution

It is the first study to systematically compare atom representations in GNNs for molecular property prediction, emphasizing their impact on model accuracy.

Findings

Atom features significantly influence GNN performance.

Different atom representations lead to varying prediction accuracies.

Featurization choices can be as important as network architecture.

Abstract

Graph neural networks have recently become a standard method for analysing chemical compounds. In the field of molecular property prediction, the emphasis is now put on designing new model architectures, and the importance of atom featurisation is oftentimes belittled. When contrasting two graph neural networks, the use of different atom features possibly leads to the incorrect attribution of the results to the network architecture. To provide a better understanding of this issue, we compare multiple atom representations for graph models and evaluate them on the prediction of free energy, solubility, and metabolic stability. To the best of our knowledge, this is the first methodological study that focuses on the relevance of atom representation to the predictive performance of graph neural networks.