Correction to: Analyzing Learned Molecular Representations for Property Prediction (J. Chem. Inf. Model. (2019) 59:8 (3370-3388) DOI: 10.1021/acs.jcim.9b00237)

Abstract

Due to an error in the processing of the random forest model's predictions on classification data sets, our original random forest AUC numbers were incorrect on six public classification data sets-HIV, BACE, BBBP, Tox21, SIDER, and ClinTox-and on one proprietary classification data set-hPXR (class). We fixed the error and reran the random forest experiments. After the fix, the random forest model performs better than previously reported, though our DMPNN continues to outperform it on some classification data sets and on all but one of the regression data sets. Additionally, since the fixed random forest model is better than our DMPNN on BACE and hPXR (class), our D-MPNN now achieves comparable or better performance than all baseline models on 11 rather than 12 of the 19 public data sets and on 15 rather than 16 of the 16 proprietary data sets. The results of the other 800+ experiments we report in the paper are unaffected. The tables and figures included here show the changes. [Figure Presented].