On Graph Neural Network Ensembles for Large-Scale Molecular Property Prediction

TL;DR

This paper presents an ensemble of three graph neural network models for large-scale molecular property prediction, achieving significant improvements over baselines and enabling uncertainty-based identification of challenging molecules.

Contribution

It introduces a novel ensemble approach combining GIN, Bayesian Neural Networks, and DiffPool for large-scale molecular property prediction.

Findings

Ensemble outperforms baseline by 7.6%

Uncertainty helps identify molecules with harder-to-predict properties

Achieves Pearson's correlation of 0.5181 for challenging molecules

Abstract

In order to advance large-scale graph machine learning, the Open Graph Benchmark Large Scale Challenge (OGB-LSC) was proposed at the KDD Cup 2021. The PCQM4M-LSC dataset defines a molecular HOMO-LUMO property prediction task on about 3.8M graphs. In this short paper, we show our current work-in-progress solution which builds an ensemble of three graph neural networks models based on GIN, Bayesian Neural Networks and DiffPool. Our approach outperforms the provided baseline by 7.6%. Moreover, using uncertainty in our ensemble's prediction, we can identify molecules whose HOMO-LUMO gaps are harder to predict (with Pearson's correlation of 0.5181). We anticipate that this will facilitate active learning.