Table 1 Comparison of DeepTMP with other state-of-the-art methods

From: Deep transfer learning for inter-chain contact predictions of transmembrane protein complexes

Method

Top 1

Top 10

Top 25

Top 50

Top L/10

Top L/5

Top L

DeepTMP

82.7 (76.9)

82.3 (76.5)

81.1 (74.8)

79.4 (71.9)

82.0 (75.3)

80.1 (72.5)

68.4 (62.3)

CDPred

55.8 (46.2)

48.5 (49.6)

48.4 (49.2)

45.9 (46.5)

48.6 (49.0)

45.6 (44.8)

33.8 (35.3)

DeepHomo2.0

50.0 (46.2)

48.7 (46.7)

47.2 (44.9)

43.3 (41.1)

44.0 (42.0)

41.1 (38.7)

31.8 (27.2)

GLINTER

44.4 (40.0)

38.0 (36.4)

38.3 (36.6)

34.3 (32.8)

39.9 (36.0)

36.4 (34.7)

27.7 (27.4)

DeepHomo

28.8 (23.1)

27.7 (23.3)

24.8 (20.5)

22.0 (19.2)

23.2 (18.7)

21.4 (18.4)

15.7 (12.1)

DNCON2_Inter

13.7 (13.7)

13.3 (13.1)

12.0 (13.0)

11.7 (13.0)

12.5 (12.8)

11.4 (12.7)

7.9 (8.3)

DeepHomo2_TMP

63.5 (61.5)

65.0 (63.3)

63.0 (59.7)

60.9 (56.2)

60.6 (57.1)

58.5 (55.9)

45.9 (40.9)

IT_Model

59.6 (57.7)

56.3 (55.8)

54.8 (53.8)

52.3 (51.0)

53.9 (52.9)

52.4 (50.9)

42.0 (39.3)

DT_Model

51.9 (55.8)

53.5 (53.3)

53.4 (53.2)

50.2 (49.0)

50.4 (51.5)

49.9 (47.9)

38.5 (37.4)

  1. The precisions (%) of DeepTMP, CDPred, DeepHomo2.0, GLINTER, DeepHomo, and DNCON2_Inter are based on the test set of 52 transmembrane protein complexes when the experimental monomer structures (predicted monomer structures by AlphaFold2) are used as input. For reference, the table also lists the results of the transfer learning model using the network architecture of DeepHomo2.0 (DeepHomo2_TMP), the initial training model (IT_Model) on the large data set of soluble protein complexes, and the direct training model (DT_Model) on the small data set of transmembrane protein complexes. The numbers in bold fonts indicate the best performances for the corresponding categories.