Skip to main content
NCBI home page
ACS AuthorChoice logoLink to ACS AuthorChoice
. 2019 Dec 9;59(12):5304–5305. doi: 10.1021/acs.jcim.9b01076

Correction to Analyzing Learned Molecular Representations for Property Prediction

Kevin Yang †,*, Kyle Swanson †,*, Wengong Jin †, Connor Coley ‡, Philipp Eiden , Hua Gao §, Angel Guzman-Perez §, Timothy Hopper §, Brian Kelley ∥, Miriam Mathea , Andrew Palmer , Volker Settels , Tommi Jaakkola †, Klavs Jensen ‡, Regina Barzilay †
PMCID: PMC8154261  PMID: 31814400

Separately, on page 3372 the learned matrix Inline graphic should be Inline graphic where Inline graphic.

Figure 1.

Figure 1

Open in a new tab

(random split, higher = better) Comparison to baselines on public data sets with original (left) and fixed (right) random forest numbers using a random split.

Table 1. (Random Split, Higher = Better) Comparison to Baselines on Public Datasets with Original and Fixed Random Forest Numbers Using a Random Split.

data set metric D-MPNN D-MPNN ensemble RF on Morgan (original) RF on Morgan (fixed)
HIV ROC-AUC 0.816 ± 0.023 0.836 ± 0.020 (+2.40% p = 0.01) 0.641 ± 0.022 (−21.45% p = 0.00) 0.819 ± 0.025 (+0.31% p = 0.97)
BACE ROC-AUC 0.878 ± 0.032 0.898 ± 0.034 (+2.31% p = 0.00) 0.825 ± 0.039 (−6.08% p = 0.00) 0.898 ± 0.031 (+2.26% p = 1.00)
BBBP ROC-AUC 0.913 ± 0.026 0.925 ± 0.036 (+1.23% p = 0.01) 0.788 ± 0.038 (−13.77% p = 0.00) 0.909 ± 0.028 (−0.42% p = 0.19)
Tox21 ROC-AUC 0.845 ± 0.015 0.861 ± 0.012 (+1.95% p = 0.00) 0.619 ± 0.015 (−26.75% p = 0.00) 0.819 ± 0.017 (−3.06% p = 0.00)
SIDER ROC-AUC 0.646 ± 0.016 0.664 ± 0.021 (+2.79% p = 0.01) 0.572 ± 0.007 (−11.38% p = 0.00) 0.687 ± 0.014 (+6.35% p = 1.00)
ClinTox ROC-AUC 0.894 ± 0.027 0.906 ± 0.043 (+1.33% p = 0.05) 0.544 ± 0.031 (−39.13% p = 0.00) 0.759 ± 0.060 (−15.12% p = 0.00)
Open in a new tab

Figure 2.

Figure 2

Open in a new tab

(scaffold split, higher = better) Comparison to baselines on public data sets with original (left) and fixed (right) random forest numbers using a scaffold split.

Table 2. (Scaffold Split, Higher = Better) Comparison to Baselines on Public Datasets with Original and Fixed Random Forest Numbers Using a Scaffold Split.

data set metric D-MPNN D-MPNN ensemble RF on Morgan (original) RF on Morgan (fixed)
HIV ROC-AUC 0.794 ± 0.016 0.817 ± 0.013 (+2.94% p = 0.00) 0.583 ± 0.034 (−26.59% p = 0.00) 0.821 ± 0.020 (+3.42% p = 0.99)
BACE ROC-AUC 0.838 ± 0.056 0.871 ± 0.041 (+3.89% p = 0.00) 0.804 ± 0.035 (−4.04% p = 0.01) 0.884 ± 0.026 (+5.43% p = 1.00)
BBBP ROC-AUC 0.888 ± 0.029 0.902 ± 0.024 (+1.56% p = 0.01) 0.722 ± 0.049 (−18.68% p = 0.00) 0.880 ± 0.034 (−0.88% p = 0.45)
Tox21 ROC-AUC 0.791 ± 0.047 0.814 ± 0.047 (+2.89% p = 0.00) 0.582 ± 0.031 (−26.42% p = 0.00) 0.747 ± 0.040 (−5.54% p = 0.00)
SIDER ROC-AUC 0.593 ± 0.032 0.612 ± 0.047 (+3.31% p = 0.03) 0.540 ± 0.013 (−8.79% p = 0.00) 0.632 ± 0.043 (+6.75% p = 1.00)
ClinTox ROC-AUC 0.870 ± 0.072 0.895 ± 0.050 (+2.86% p = 0.01) numerically unstable 0.711 ± 0.123 (−18.24% p = 0.00)
Open in a new tab

Figure 3.

Figure 3

Open in a new tab

(time split, higher = better) Comparison to baselines on Amgen data set with original (left) and fixed (right) random forest numbers using a time split.

Table 3. (Time Split, Higher = Better) Comparison to Baselines on Amgen Dataset with Original and Fixed Random Forest Numbers Using a Time Split.

data set metric D-MPNN D-MPNN ensemble RF on Morgan (original) RF on Morgan (fixed)
hPXR (class) ROC-AUC 0.842 ± 0.008 0.858 ± 0.002 (+1.95%) 0.598 ± 0.004 (−28.98%) 0.869 ± 0.007 (+3.28%)
Open in a new tab

Table 4. Number of Public Datasets Where D-MPNN is Statistically Significantly Better than, Equivalent to, or Worse than Random Forest.

baseline D-MPNN is better D-MPNN is the same D-MPNN is worse no. data sets
RF on Morgan (original) 14 0 1 15
RF on Morgan (fixed) 9 1 4 15
Open in a new tab

Articles from Journal of Chemical Information and Modeling are provided here courtesy of American Chemical Society

RESOURCES