Skip to main content

Table 1 Performance statistics for the models described here and their beta binomial parameters a

From: Using beta binomials to estimate classification uncertainty for ensemble models

Data set

Model

Architect.b

Threshold

Sensitivity

Specificity

J c

α

β

MRd

logP2

1

33×6×40

16.5

0.849

0.882

0.731

0.695c

0.772

0.083

 

2

33×3×45

16.5

0.861

0.857

0.718

0.635

0.571

0.087

Ames

1

33×2×26

16.5

0.791

0.596

0.387

0.926

0.537

0.222

 

2

33×4×24

16.5

0.700

0.676

0.376

0.489

0.462

0.239

logP3

1a

33×4×44

16.5

0.882

0.892

0.774

1.229

0.469

0.077

 

1b

33×4×44

24.5

0.840

0.921

0.761

0.925

0.323

0.066

 

2a

75×4×42

37.5

0.889

0.885

0.775

1.163

0.472

0.089

 

2b

75x4x42

53.5

0.858

0.910

0.768

1.037

0.415

0.083

CYP2D6

1a

33×3×35

16.5

0.721

0.789

0.510

1.561

0.447

0.211

 

1b

33×3×35

27.5

0.604

0.873

0.476

1.350

0.294

0.164

logP3

3a e

33×3×24 e

16.2

0.874

0.862

0.736

0.891

0.263

0.095

 

3b

33×3×24

20.3

0.862

0.874

0.742

0.690

0.306

0.096

  1. aPerformance statistics are for compounds in the validation set. Beta binomial parameters shown are those obtained by fitting to the distribution of training pool misclassifications. b“Architect”. indicates the network architecture, given as networks × neurons × inputs. cJ is Youden’s index [28], which is equal to sensitivity + specificity – 1. d“MR” is the misclassification rate for the training pool. eBoldface names and architectures indicate models in which ensemble classifications were determined using the averaging method.