Table 1 Comparison of generated molecules from our procedure (DIDgen) with those generated with a state-of-the-art genetic algorithm (JANUS)35

From: Using GNN property predictors as molecule generators

Method

4.1 eV

6.8 eV

9.3 eV

ρ

 

ncalcs

n±0.5

MAE

Div.

ncalcs

n±0.5

MAE

Div.

ncalcs

n±0.5

MAE

Div.

 

QM9

 

3.4%

 

0.87

 

24%

 

0.89

 

6.0%

 

0.84

–

JANUS (DFT)

197

24 (12.2%)

0.96

0.79

392

42 (10.7%)

0.92

0.80

484

26 (5.4%)

1.28

0.81

–

JANUS (Proxy)

100

36

1.05

0.86

100

46

0.80

0.82

100

37

1.24

0.81

0.78

DIDgen

100

46

0.81

0.91

100

50

0.83

0.90

100

34

0.83

0.83

0.86

  1. The comparison metrics are the number of density functional theory (DFT) calculations (ncalcs), the number of molecules that are within 0.5 eV of the target (n±0.5), the mean absolute error (MAE) to the target property and the diversity (Div.) of the generated molecules that are within 0.5 eV of the property. Pearson correlation coefficients (ρ) between all 300 ML and DFT predictions are reported in the last column. The highest performance for each metric is in bold.
Back to article page