Table 2 Performance of different GNN architectures for regression tasks

From: Geometry-enhanced molecular representation learning for property prediction

 

RMSE

MAE

Method

ESOL

FreeSolv

Lipo

QM7

QM8

QM9

GIN19

1.067(0.051)

2.346(0.122)

0.757(0.022)

110.3(7.2)

0.0199(0.0002)

0.00886(0.00005)

GAT46

1.556(0.085)

3.559(0.050)

1.021(0.029)

103.0(4.4)

0.0224(0.0005)

0.01117(0.00018)

GCN47

1.211(0.052)

3.174(0.308)

0.773(0.007)

100.0(3.8)

0.0203(0.0005)

0.00923(0.00019)

D-MPNN43

1.050(0.008)

2.082(0.082)

0.683(0.016)

103.5(8.6)

a0.0190(0.0001)

0.00814(0.00009)

AttentiveFP44

a0.877(0.029)

a2.073(0.183)

a0.721(0.001)

a72.0(2.7)

0.0179(0.0001)

a0.00812(0.00001)

GTransformer4

2.298(0.118)

4.480(0.155)

1.112(0.029)

161.3(7.1)

0.0361(0.0008)

0.00923(0.00019)

SGCN16

1.629(0.001)

2.363(0.050)

1.021(0.013)

131.3(11.6)

0.0285(0.0005)

0.01459(0.00055)

DimeNet17

0.878(0.023)

2.094(0.118)

0.727(0.019)

95.6(4.1)

0.0215(0.0003)

0.01031(0.00076)

HMGNN6

1.39(0.073)

2.123(0.179)

2.116(0.473)

101.6(3.2)

0.0249(0.0004)

0.01239(0.0001)

GeoGNN

0.832(0.010)

1.857(0.071)

0.666(0.015)

59.0(3.4)

0.0173(0.0004)

0.00746(0.00003)

  1. The SOTA results are shown in bold.
  2. aThe cells in grey indicate the previous SOTA results.
Back to article page