Skip to main content

Table 4 The performance comparison for classification and regression tasks

From: ABT-MPNN: an atom-bond transformer-based message-passing neural network for molecular property prediction

Classification (the higher the better)a

 

Johnson et al

Tox21

Clintox

ToxCast

HIV

RF

0.252 \(\pm\) 0.014

0.818 \(\pm\) 0.005

0.721 \(\pm\) 0.088

_b

0.798 \(\pm\) 0.040

FFN

0.258 \(\pm\) 0.015

0.837 \(\pm\) 0.010

0.837 \(\pm\) 0.062

0.738 \(\pm\) 0.009

0.803 \(\pm\) 0.045

MPNN

0.258 \(\pm\) 0.013

0.859 \(\pm\) 0.011

0.873 \(\pm\) 0.051

0.752 \(\pm\) 0.010

0.788 \(\pm\) 0.050

D-MPNN

0.281 \(\pm\) 0.028

0.855 \(\pm\) 0.015

0.895 \(\pm\) 0.037

0.749 \(\pm\) 0.013

0.788 \(\pm\) 0.039

Deeper GCN

0.272 \(\pm\) 0.022

0.853 \(\pm\) 0.013

0.870 \(\pm\) 0.042

0.751 \(\pm\) 0.010

0.789 \(\pm\) 0.031

GEM

0.280 \(\pm\) 0.018

0.864 \(\pm\) 0.010

0.825 \(\pm\) 0.091

0.757 \(\pm\) 0.013

0.769 \(\pm\) 0.038

ABT-MPNN

0.295 \(\pm\) 0.021

0.857 \(\pm\) 0.010

0.904 \(\pm\) 0.034

0.760 \(\pm\) 0.013

0.809 \(\pm\) 0.036

Regression (the lower the better)a

 

Johnson et al

ESOL

Lipophilicity

Freesolv

QM8

RF

1.315 \(\pm\) 0.021

1.230 \(\pm\) 0.066

0.846 \(\pm\) 0.039

2.467 \(\pm\) 0.570

0.014 \(\pm\) 0.000

FFN

1.321 \(\pm\) 0.016

0.614 \(\pm\) 0.109

0.674 \(\pm\) 0.043

1.275 \(\pm\) 0.352

0.016 \(\pm\) 0.000

MPNN

1.309 \(\pm\) 0.017

0.575 \(\pm\) 0.086

0.585 \(\pm\) 0.044

1.042 \(\pm\) 0.220

0.010 \(\pm\) 0.000

D-MPNN

1.307 \(\pm\) 0.024

0.594 \(\pm\) 0.066

0.558 \(\pm\) 0.044

0.915 \(\pm\) 0.142

0.010 \(\pm\) 0.000

Deeper GCN

1.325 \(\pm\) 0.015

0.601 \(\pm\) 0.056

0.580 \(\pm\) 0.035

0.970 \(\pm\) 0.368

0.012 \(\pm\) 0.000

GEM

1.315 \(\pm\) 0.021

0.632 \(\pm\) 0.062

0.599 \(\pm\) 0.035

0.962 \(\pm\) 0.257

0.010 \(\pm\) 0.000

ABT-MPNN

1.305 \(\pm\) 0.017

0.566 \(\pm\) 0.075

0.554 \(\pm\) 0.041

0.902 \(\pm\) 0.157

0.009 \(\pm\) 0.000

  1. aThe evaluation metrics are represented as averaged values ± standard deviation from fivefold CV. The best performance values are highlighted in bold
  2. bThe results of RF on ToxCast are not presented because of the substantial computational cost