Skip to main content

Table 1 Metrics calculated on the top 500 high-scoring molecules obtained from a multi-objective molecule generation where DRD2 bioactivity, hERG inactivity, and QED were optimized simultaneously (†Total Oracle score combines both DRD2 bioactivity and hERG inactivity oracles, as well as the QED score via Eq. 1. RO3 MolLogP is a lead-like filter which corresponds to the percentage of molecules that satisfy the condition MolLogP \(< 3\))

From: Human-in-the-loop active learning for goal-oriented molecule generation

Metric (mean)

No feedback

With feedback on DRD2 bioactivity

Chemist 1

Chemist 2

Chemist 3

DRD2 Predicted score

0.93

0.80 **

0.81 **

0.84 **

DRD2 Oracle score \(\uparrow\)

0.50

0.55

0.49

0.74 **

Mean Absolute Error \(\downarrow\)

0.42

0.25 **

0.32 **

0.10 **

QED score \(\uparrow\)

0.57

0.61 **

0.58

0.71 **

hERG inactivity score \(\uparrow\)

0.91

0.88

0.90

0.82

Total Oracle score† \(\uparrow\)

0.68

0.69

0.67

0.77 **

Internal Diversity \(\uparrow\)

0.47

0.41

0.45

0.44

RO3 MolLogP \(\uparrow\)

0.70

0.54 **

0.79 **

0.66

SA score \(\downarrow\)

3.04

2.75 **

2.82 **

3.08

Novelty \(\uparrow\)

1.0

1.0

1.0

1.0

Uniqueness \(\uparrow\)

1.0

1.0

1.0

1 .0

  1. Values that significantly differ from the "No Feedback" baseline (ANOVA p-value < 0.01) are marked with the superscript **. Values in bold correspond to the most performant methods in comparison with the "No Feedback" baseline