Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models

Abstract Graph neural networks (GNN) has been considered as an attractive modelling method for molecular property prediction, and numerous studies have shown that GNN could yield more promising results than traditional descriptor-based methods. In this study, based on 11 public datasets covering various endpoints, the predictive capacity computational efficiency of prediction models developed by eight machine learning (ML) algorithms, including four (SVM, XGBoost, RF DNN) graph-based (GCN, GAT, MPNN Attentive FP), were extensively tested compared. The demonstrate average outperform in terms accuracy efficiency. SVM generally achieves best predictions regression tasks. Both XGBoost can achieve reliable classification tasks, some models, such FP GCN, outstanding performance a fraction larger or multi-task datasets. cost, are two most efficient algorithms only need few seconds to train model even large dataset. interpretations SHAP effectively explore established domain knowledge models. Finally, we explored use these virtual screening (VS) towards HIV demonstrated different ML offer diverse VS profiles. All all, believe off-the-shelf still be directly employed accurately predict chemical endpoints with excellent computability interpretability.