Skip to main content

Table 2 Effect of model architecture and loss function choice on FoldHSphere performance using the LINDAHL dataset

From: FoldHSphere: deep hyperspherical embeddings for protein fold recognition

Model

Family

Superfamily

Fold

Top 1

Top 5

Top 1

Top 5

Top 1

Top 5

(a) Softmax loss

CNN-GRU [48]

68.6 (1.94)

89.2 (1.37)

56.2 (2.34)

77.4 (1.96)

56.7 (2.82)

74.1 (2.46)

CNN-BGRU [48]

71.0 (1.92)

87.7 (1.42)

60.1 (2.30)

77.2 (2.02)

58.3 (2.83)

78.8 (2.27)

ResCNN-GRU

72.6 (1.87)

90.3 (1.24)

59.4 (2.32)

77.0 (2.00)

58.9 (2.88)

75.1 (2.44)

ResCNN-BGRU

76.8 (1.78)

91.2 (1.23)

65.0 (2.29)

82.0 (1.84)

59.5 (2.79)

76.6 (2.35)

(b) LMCL

CNN-GRU

76.6 (1.80)

90.8 (1.25)

64.7 (2.21)

80.2 (1.90)

65.7 (2.69)

79.8 (2.22)

CNN-BGRU

76.2 (1.79)

89.4 (1.31)

70.5 (2.12)

83.2 (1.80)

72.0 (2.48)

81.0 (2.21)

ResCNN-GRU

75.7 (1.77)

89.7 (1.25)

66.4 (2.29)

81.1 (1.86)

67.6 (2.63)

80.1 (2.23)

ResCNN-BGRU

75.1 (1.84)

89.5 (1.30)

69.8 (2.25)

85.3 (1.67)

74.1 (2.42)

82.2 (2.12)

(c) Thomson LMCL

CNN-GRU

80.0 (1.73)

90.6 (1.24)

66.8 (2.23)

80.2 (1.94)

66.0 (2.62)

80.1 (2.22)

CNN-BGRU

77.5 (1.75)

91.7 (1.19)

69.8 (2.09)

85.3 (1.64)

72.6 (2.48)

82.2 (2.14)

ResCNN-GRU

76.9 (1.78)

89.5 (1.28)

69.1 (2.20)

82.9 (1.77)

69.5 (2.57)

79.4 (2.26)

ResCNN-BGRU

76.4 (1.77)

89.2 (1.30)

72.8 (2.15)

86.4 (1.63)

75.1 (2.47)

84.1 (2.12)

  1. The fold recognition accuracy (%) results are provided at the family, superfamily and fold levels, considering both the top 1 and top 5 ranked templates. We compare the CNN-GRU, CNN-BGRU, ResCNN-GRU and ResCNN-BGRU neural network models, trained with different loss functions: (a) Softmax loss with sigmoid activation, (b) LMCL with tanh activation, and (c) Thomson LMCL with tanh activation. Optimal LMCL hyperparameters are in Table 1. Boldface indicates the best performance per loss function. For each accuracy result, we also provide in parentheses the standard error estimated using 1000 bootstraps