Skip to main content

Table 10 The mean wall-clock time (seconds) for the six single-task datasets given by the four descriptor-based and four graph-based models

From: Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models

Dataset

SVMa

XGBoostb

RFb

DNNc

GCNd

GATd

MPNNd

Attentive FPd

FreeSolv (639)

0.17

0.209

1.429

6.27

18.458

29.37

77.85

20.927

ESOL (1127)

0.51

0.329

0.342

9.032

68.197

80.597

181.114

59.199

Lipop (4200)

6.431

7.379

5.722

28.686

159.879

151.191

611.048

652.777

BACE (1513)

2.105

0.327

1.327

8.911

108.967

156.074

630.748

137.291

BBBP (2035)

8.033

0.242

0.873

6.74

83.062

129.817

316.224

98.743

HIV (40748)

852.312

23.653

14.118

215.965

867.148

1122.126

1867.602

677.536

  1. aSVM was implemented with the scikit-learn package and run in a single thread (CPU: Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10 GHz); bXGBoost and RF were implemented with the scikit-learn package and run in six parallel threads (CPU: Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10 GHz); cDNN was implemented with PyTorch package and run in a single GPU card (NVIDIA GEFORCE RTX 2080 Ti with video memory of 11G); dGCN, GAT, MPNN and Attentive FP were implemented with DGL package using PyTorch as the backend and run in a single GPU card (NVIDIA GEFORCE RTX 2080 Ti with video memory of 11G); All tested NN-based models were trained with a batch-size 128 in early-stopping way as described in ‘Materials and methods’ (HIV with a batch-size 128*5 due to the large data volume)