truPK

Technology Overview

Prediction Models for Human PK

The use of in silico prediction of ADME/Tox properties is gaining acceptance as a useful assessment tool for early identification of likely drug candidate failures. However, until now, it has been difficult to locate reliable models for the prediction of human pharmacokinetics in silico.

Strand Genomics Pvt. Ltd. has developed five human PK models and is offering them either as trupk or in partnership with Bio-Rad Informatics as part of the KnowItAll Informatics System, ADME/Tox Edition.

Several machine-learning methods including neural networks, decision trees and support vector machines were employed to identify a small set from 1054 molecular descriptors that correlated with each pharmacokinetic parameter. The input to the predictors is the 2-D structure of a molecule, which is used to compute the descriptors that are utilized by the models.

Human PK Models

Model	Units	Comments
Plasma Protein Binding	Fraction of total drug	Total drug in the plasma of drug bound to the plasma proteins
Bioavailability	Fraction of total oral dose	Total oral dose of a drug that reaches the plasma upon absorption in the gut and first pass metabolism.
Volume of Distribution	L/Kg	Measure of the total distribution volume available to a drug in the body. Low volumes of distribution imply that the drug remains in the plasma while very high volumes indicate that the drug distributes widely into the various tissue compartments in the body.
Elimination Half-life	Hours	Measures the half-life of a drug observed in the plasma.
Rate of Absorption	1/hour	Measures the rate at which a drug is absorbed orally from the gut into the plasma. Quickly absorbed drugs will exhibit a very fast time to maximum concentration in the plasma.

Training Set Profile

The activity values used in the generation of the Strand human PK models were taken from the following sources:

FDA CDER Human Drugs Database
RxList Drug Information Guide
Thummel K.E., Shen D.D., in "Goodman & Gilman's The Pharmacological Basis of Therapeutics", 10/e, Ed. by Hardman J.G., Limbird L.E., Gilman A.G., McGraw-Hill, pp. 1924-2023

Number of Lipinski’s Rule Violations	Number of Molecules
0	378
1	49
2	30
3	17
4	0

All the molecules used in building the models are commercially available drugs. 90% of the drugs listed here are considered drug-like based on Lipinski’s classification. The detailed list is on the left-

Average number of H-bond donors per compound: 2.26
Average number of H-bond acceptors per compound: 6.2
Average molecular weight: 351
Percent of drug like compounds: 90%

A self-dissimilarity test for the training set was carried out by using 2D structural fingerprints and was determined to be 52%. The maximal dissimilarity was observed to be 93 %

Model Characteristics: Training and External Validation

Training Statistics:

		Classification Accuracy		Regression Accuracy
	N	Low %	High %	% Accurately Predicted	R-squared
Protein Binding	306	99	100	85	0.93
Bioavailability	185	--	--	89	0.80
Volume of Distribution	206	99	87	97	0.86
Elimination Half life	341	100	100	82	0.82
Rate of Absorption	193	89	86	97	0.86

Cross Validation Statistics:

		Classification Accuracy		Regression Accuracy
	Validation Type	Low %	High %	% Accurately Predicted	Robust Q-squared
Protein Binding	N-Fold	75	75	78	0.87
Bioavailability	N-Fold	--	--	75	0.65
Volume of Distribution	N-Fold	69	67	89	0.73
Elimination Half life	N-Fold	76	85	73	0.67
Rate of Absorption	N-Fold	78	76	93	0.78

External Validation Statistics:

		Classification Accuracy		Regression Accuracy
	N	Low %	High %	% Accurately Predicted	R-squared
Protein Binding	74	82	97	75	0.90
Bioavailability	42	--	--	86	0.70
Volume of Distribution	53	92	86	79	0.73
Elimination Half life	66	85	65	71	0.72
Rate of Absorption	30	50	90	63	0.67