Table 3 Classification performance of different fine-tuning modes.

From: Hierarchical Molecular Graph Self-Supervised Learning for property prediction

Methods

Version

BACE

BBBP

Tox21

ToxCast

SIDER

ClinTox

GIN

SMALL

72.6

70.2

72.2

62.7

57.5

60.7

 

LARGE

73.4

67.7

73.8

62.5

56.2

60.8

HiMol (no pre-train)

SMALL

83.3

67.5

74.1

64.4

58.7

60.0

 

LARGE

79.6

68.4

74.7

63.2

62.1

62.6

HiMol (freeze)

SMALL

77.9

59.2

68.1

60.9

61.1

66.0

 

LARGE

72.6

60.3

65.4

59.3

61.5

67.6

HiMol

SMALLa

84.6

71.3

76.0

66.0

62.5

70.6

 

LARGEb

84.3

73.2

76.2

66.3

61.3

80.8

  1. aSMALL version implements 3-layer GIN as the GNN backbone.
  2. bLARGE version implements 5-layer GIN as the GNN backbone.
  3. The values in bold highlight the best performing results of each benchmark.
Back to article page