Computational Drug Discovery and Design

3.4 Performance
Evaluation

Various threshold dependent and threshold independent perfor-

mance evaluation metrics can be used for judging the performance

of the machine learning algorithms (seeNote 3).

Sensitivity: This can be defined as % of correctly predicted drug

targets.

Sensitivity¼

TP

ðÞTPþFN

 100 ð 1 Þ

Specificity: This can be defined as % of correctly predicted

nondrug targets.

Specificity¼

TN

ðÞTNþFP

 100 ð 2 Þ

Accuracy: This can be defined as the % of correctly predicted

drug targets and nondrug targets.

Accuracy¼

TPþTN




TPþFPþTNþFN

 100 ð 3 Þ

Matthews Correlation Coefficient (MCC):For binary classi-

fication problems it’s a useful performance evaluation metric. Its

values ranges from1 to +1 (worse to best).

MCC¼

ðÞTPTN ðÞFPFN

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

ðÞTPþFNðÞTPþFPðÞTNþFPðÞTNþFN

p ð 4 Þ

Youden’s index:This performance evaluation metric gives an

indication about the model’s ability to avoid failures. Higher values

are better.

Y¼SensitivityðÞð 1 Specificity 5 Þ

Area under the Curve (AUC):The area under the receiver

operation characteristic curves know as AUC and can be used to

summarize the ROC by a single numerical quantity. Its values

ranges from 0 to 1 and is threshold independent [39].

g -means: This is the geometric mean of sensitivity and

specificity

gmeans¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

SensitivtySpecificity

p

ð 6 Þ

3.5 Conclusion
and Future Perspective

Machine learning methods have advantage over sequence align-

ment based methods as they can take into account of the hidden

similarities between features for generating successful prediction

models. Sequence feature generation step should account to cover

as much as possible of chemical and genomic space. Protein–pro-

tein interaction data notably from databases like STRING [40],

BioGRID [41] and Human Protein Reference Databases(HPRD)

Human Drug Targets and Their Interactions 27