Interpretable Molecular Property Predictions Using Marginalized Graph Kernels

Yan Xiang

doi:10.1021/acs.jcim.3c00396

Interpretable Molecular Property Predictions Using Marginalized Graph Kernels

2023-07-28

3

PRE

AI

DOI：10.1021/acs.jcim.3c00396

Original

Original request for help

Save

Abstract

En 中文

Marginalized graph kernels have shown competitive performanceinmolecular machine learning tasks but currently lack measures of interpretability,which are important to improve trust in the models, detect biases,and inform molecular optimization campaigns. We here conceive andimplement two interpretability measures for Gaussian process regressionusing a marginalized graph kernel (GPR-MGK) to quantify (1) the contributionof specific training data to the prediction and (2) the contributionof specific nodes of the graph to the prediction. We demonstrate theapplicability of these interpretability measures for molecular propertyprediction. We compare GPR-MGK to graph neural networks on four logicand two real-world toxicology data sets and find that the atomic attributionof GPR-MGK generally outperforms the atomic attribution of graph neuralnetworks. We also perform a detailed molecular attribution analysisusing the FreeSolv data set, showing how molecules in the trainingset influence machine learning predictions and why Morgan fingerprintsperform poorly on this data set. This is the first systematic examinationof the interpretability of GPR-MGK and thereby is an important stepin the further maturation of marginalized graph kernel methods forinterpretable molecular predictions.

Keywords:

STRUCTURAL ALERTS

Interpretable Molecular Property Predictions Using Marginalized Graph Kernels

Abstract

Journal

Organization