truTox

Technology Overview

Prediction model for HERG binding

QT elongation caused by drugs binding to the hERG channel can result in potentially fatal cardiac arrhythmias. The use of in silico models to predict the hERG binding potential of a drug (a surrogate marker for cardiotoxicity) has been gaining acceptance as a means of flagging candidates that could potentially pose problems in advanced stages of clinical development. Such models can also provide rational guidelines for molecule design to overcome this problem.

Strand Genomics has developed a suite of hERG binding predictors employing several machine learning methods such as neural networks and support vector machines to identify a small set of molecular descriptors that correlates with hERG binding activity. The input to these predictors is the 2D structure of a molecule, which is used to compute the descriptors.

Binary Classification Model for HERG Binding

hERG Binding activity is expressed as IC₅₀, which is the concentration of drug required to inhibit I_kr by 50%. Typically binders and nonbinders to hERG are defined by the selecting an appropriate IC₅₀ cutoff value. A high value for this cutoff implies that the model is stringent in classifying compounds as binders. One can relax the stringency by using lower cutoff values. Strand has developed three models for hERG binding that differ in the IC₅₀ cutoff values. These models in increasing order of stringency are
a) Model 1: cutoff value is IC₅₀ 1µm
b) Model 2: cutoff value is IC₅₀ 5µm
c) Model 3: cutoff value is IC₅₀ 50µm
The flexibility in choosing models of different stringencies allows the user to pick the model most applicable for decision-making. For example, a user may wish to use a model with lower stringency at a very early design stage while using a more stringent model at a later stage of development. Only Strand offers this flexibility to the user.

Training Set Profile

Data for the training set was curated from multiple public domain sources and normalized to a common basis prior to model building. Models 1 through 3 contain training sets with experimental hERG binding IC₅₀ values for 127, 170 and 204 compounds respectively.

Almost all molecules used in the model are commercially available drugs. 98% of these drugs are considered drug-like based on Lipinski's rule of five.

Number of Lipinski’s Rule Violations	Number of Molecules
0	170
1	28
2	3
3	3
4	0

A self-dissimilarity test for the training set was carried by using 2D structural fingerprints and was determined to be 64%. The maximal dissimilarity was observed to be 96%.

Model Characteristics: Validation

Cross Validation Statistics

IC₅₀ cutoff (µm)	Number of Compounds	Model Type	Classification Accuracy
IC₅₀ cutoff (µm)	Number of Compounds	Model Type	Binder	Non Binder	Overall
1	127	ANN	93%	81%	86%
5	170	ANN	84%	86%	85%
50	204	ANN	77%	78%	78%
50	204	SVM	80%	77%	78%

External Validation Statistics

IC₅₀ cutoff (µm)	Number of Compounds	Model Type	Prediction Accuracy
IC₅₀ cutoff (µm)	Number of Compounds	Model Type	Binder	Non Binder	Overall
1	199	ANN	100%	80%	81%
5	191	ANN	100%	77%	78%
50	189	ANN	88%	76%	77%
50	189	SVM	83%	79%	80%