Table 1 Time-based PDBBind docking accuracy benchmark (n = 333)

From: Interformer: an interaction-aware model for protein-ligand docking and affinity prediction

Models

TOP-1 RMSD

TOP-5 RMSD

Time (s)↓

 

% < 2Å↑

Median↓

% < 2Å↑

Median↓

 

Blind docking

 EquiBind3

5.5

6.2

–

–

0.04

 TANKBind39

20.4

4.0

24.5

3.4

2.5

 E3Bind4

23.4

3.8

–

–

2.2

 DiffDock5

38.2

3.3

44.7

2.4

40

Pocket residues specified

 AutoDock Vina17

44.6

2.72

62.6

1.42

41.2

 GNINA18

53.3

1.79

68.8

1.27

194.2

 DeepDock16

18.8

4.80

–

–

64.3

Our Interformer

 Energy

63.9

1.25

78.0

1.02

31.9

 w/ Pose Score

62.1

1.38

77.17

0.99

32.9

  1. The percentage of predictions with RMSD less than 2 Å, the median RMSD, and the method average runtime are shown. Energy indicates the use of an energy model for sampling docking pose, w/ Pose Score indicates the sampled docking pose is further ranked by the pose score models. The performance results for EquiBind, TANKBind, and DiffDock are obtained from the DiffDock paper5. The results for E3Bind is obtained from its paper4. AutoDock Vina, GNINA and DeepDock are re-performed from this work. Bold-formatted numbers represent the best-performing result on each performance metric.
Back to article page