Skip to main content

Table 5 Target Family Performance Comparison

From: Industry-scale application and evaluation of deep learning for drug target prediction

 

ExCAPE-ML

AstraZeneca

Janssen

Target family

Targets w./sz.

p-value

Target w./sz.

p-value

Targets w./sz.

p-value

Oxidoreductase

25/37

2.17e05

7/18

3.91e−01

16/32

3.77e−02

Transferase

141/158

1.58e48

77/143

3.64e−07

60/153

7.38e−02

Hydrolase

63/92

6.24e12

21/47

6.96e−02

36/77

9.91e03

Lyase

4/8

2.59e−01

0/0

 

4/8

2.59e−01

Isomerase

4/6

1.00e−01

0/1

1.00e + 00

4/6

1.00e−01

GPCR Fam. A

76/94

3.78e21

34/70

5.87e03

66/93

1.39e13

GPCR Fam. B

3/5

2.10e−01

0/5

1.00e + 00

2/5

5.39e-01

GPCR Fam. C

3/5

2.10e−01

1/5

8.68e-01

5/5

4.12e-03

Nuclear Hormone Receptor

14/20

8.79e04

9/17

7.55e−02

13/19

1.87e03

Reader

2/7

7.37e−01

1/1

3.33e−01

2/2

1.11e−01

Eraser

7/9

8.28e−03

5/7

4.53e-02

4/7

1.73e-01

Writer

3/3

3.70e−02

1/1

3.33e−01

0/1

1.00e + 00

Ligand-gated

3/6

3.20e−01

1/4

8.02e−01

2/6

6.49e−01

Voltage-gated

6/12

1.78e−01

3/9

6.23e−01

6/12

1.78e−01

Primary active

3/4

1.11e−01

0/2

1.00e + 00

2/4

4.07e−01

Electrochem.

7/10

1.97e−02

2/8

8.05e−01

3/8

5.32e−01

Overall

364/476

1.26e82

162/338

2.02e08

225/438

5.95e15

  1. Number of targets won (w.) by DNNs from a target family, size of target family (sz.) and p-values of binomial tests for each target family class, with the null hypothesis that the probability of being the best method for a certain target is less than 1/3 for DNNs when compared to XGB and MF
  2. p-values below the significance threshold of 0.01 are in italics