Table 1 Overall performance comparison to the state-of-the-art methods on molecular property prediction classification tasks.

From: Pharmacophoric-constrained heterogeneous graph transformer model for molecular property prediction

Classification (ROC-AUC%, higher is better↑)

Dataset

BBBP

BACE

ClinTox

Tox21

SIDER

HIV

Molecules

2039

1513

1478

7831

1427

41127

Task

1

1

2

12

27

1

Splitting strategy

Scaffold

Scaffold

Scaffold

Scaffold

Scaffold

Scaffold

AttentiveFP

90.8 (5.01)

78.4 (0.02)

93.3 (2.04)

80.7 (2.02)

60.5 (6.01)

75.7 (1.40)

FragGAT

92.3 (4.04)

80.1 (0.86)

93.9 (2.06)

82.3 (1.62)

63.3 (3.23)

76.1 (0.65)

MGSSL

69.7 (0.91)

-

80.7 (2.12)

76.5 (0.31)

61.8 (0.81)

-

MPNN

91.3 (4.14)

77.9 (1.62)

87.9 (5.25)

80.8 (2.39)

59.5 (3.03)

74.1 (1.15)

DMPNN

91.9 (3.04)

80.9 (0.60)

89.7 (4.01)

82.6 (2.32)

63.2 (2.28)

78.6 (1.40)

CMPNN

92.7 (0.23)

82.1 (0.64)

90.2 (1.20)

80.6 (1.57)

61.6 (0.31)

77.4 (0.50)

CoMPT

93.8 (2.13)

81.9 (1.26)

93.4 (1.85)

80.9 (1.40)

63.4 (2.97)

78.1 (2.60)

GROVERbase

93.6 (0.80)

82.6 (0.70)

92.5 (1.30)

81.9 (2.00)

65.6 (0.60)

62.5 (0.90)

GROVERlarge

94.0 (1.90)

81.0 (1.40)

94.4 (2.10)

83.1 (2.50)

65.8 (2.30)

68.2 (1.10)

PharmHGT

95.4 (1.15)

86.5 (2.21)

94.5 (0.42)

83.9 (0.56)

66.9 (1.63)

80.6 (0.21)

  1. The results of baselines are obtained by us using a 5-fold cross-validation with scaffold split and doing experiment on each task for one time. The values in this table are the Mean and standard deviation of ROC-AUC values. The best performance is marked in bold and the second best is underlined to facilitate reading.