Table 1 The average performance and standard deviation of various methods on the new 115-endpoint acute toxicity dataset via 5-fold cross-validation

From: ToxACoL: an endpoint-aware and task-focused compound representation learning paradigm for acute toxicity assessment

Methods

All 115 endpoints

68 Small-sized endpoints

11 Human endpoints

 

R2 ↑

RMSE ↓

R2 ↑

RMSE ↓

R2 ↑

RMSE ↓

ST-DNN

0.11 ± 0.47

1.10 ± 0.38

−0.13 ± 0.45

1.28 ± 0.34

−0.50 ± 0.24

1.70 ± 0.32

ST-RF

0.21 ± 0.42

1.01 ± 0.35

0.01 ± 0.42

1.16 ± 0.32

−0.39 ± 0.25

1.60 ± 0.30

GAT

0.08 ± 0.32

0.94 ± 0.26

−0.01 ± 0.38

1.00 ± 0.30

−0.06 ± 0.24

1.23 ± 0.26

GCN

0.12 ± 0.43

0.91 ± 0.29

−0.04 ± 0.48

1.00 ± 0.32

−0.01 ± 0.19

1.19 ± 0.21

Attentive FP

0.32 ± 0.22

0.82 ± 0.24

0.26 ± 0.25

0.87 ± 0.27

0.24 ± 0.12

1.06 ± 0.24

MT-DNN

0.36 ± 0.37

0.77 ± 0.32

0.22 ± 0.40

0.88 ± 0.34

−0.22 ± 0.23

1.34  ± 0.30

MT-GCN

0.34 ± 0.35

0.80 ± 0.32

0.20 ± 0.38

0.91 ± 0.33

−0.25 ± 0.20

1.34 ± 0.31

DLCA

0.39 ± 0.32

0.76 ± 0.29

0.28 ± 0.35

0.86 ± 0.32

−0.15 ± 0.18

1.28 ± 0.29

ToxACoL

0.51 ± 0.24

0.68 ± 0.23

0.44 ± 0.28

0.73 ± 0.26

0.28 ± 0.24

1.02 ± 0.26

  1. A small-sized endpoint contains no more than 200 available toxicity measurement data. An upward arrow next to each metric indicates that higher values represent better performance, while the upward arrow indicates the opposite. Best performance among all methods for each metric is shown in bold.
Back to article page