ADME evaluation: 2. A computer model for the prediction of intestinal absorption in humans

https://doi.org/10.1016/S0928-0987(02)00219-1Get rights and content

Abstract

Purpose: To develop a computational method to rapidly evaluate human intestinal absorption, one of the drug properties included in the term ADME (Absorption, Distribution, Metabolism, Excretion). Poor ADME properties are the most important reason for drug failure in clinical development. Methods: The model developed is based on a modified contribution group method in which the basic parameters are structural descriptors identified by the case program, together with the number of hydrogen bond donors. Results: The human intestinal absorption model is a quantitative structure–activity relationship (QSAR) that includes 37 structural descriptors derived from the chemical structures of a data set containing 417 drugs. The model was able to predict the percentage of drug absorbed from the gastrointestinal tract with an r2 of 0.79 and a standard deviation of 12.32% of the compounds from the training set. The standard deviation for an external test set (50 drugs) was 12.34%. Conclusions: The availability of reliable and fast models like the one we propose here to predict ADME/Tox properties could help speed up the process of finding compounds with improved properties, ultimately making the entire drug discovery process shorter and more cost efficient.

Introduction

The successful development of a new drug depends on a number of criteria that have to be met. The most important, of course, is that it displays substantial beneficial activity in the treatment of a particular disease. This implies that in addition to intrinsic activity, the drug is able to reach its target and does not produce overwhelming toxic effects. Many active drugs fail in phase II or III of the development process because they do not reach their intended target. The development of tools capable to predict the factors linked to drug availability at the target site, or oral bioavailability, would therefore greatly enhance our ability to develop viable drugs. These factors are seen as depending on the drugs rate and extent of Absorption into the bloodstream, Distribution and binding to tissues, as well as rate and extent of Metabolism and Excretion or ‘ADME’.

We have previously published a paper relating our efforts to create a model for protein binding (Klopman et al., 2000) based on the use of mcase, an artificial intelligence program capable of automatically discovering the relation, if it exists, between the molecular structure of a set of chemicals and their biological activity (Klopman, 1992). This article is the second of our papers about the evaluation of ADME properties using computer-based models, and deals mainly with intestinal permeability or absorption.

Recently, Wessel and Mente (2001) published an excellent review on the prediction of ADME by computer. The authors reviewed the major methods and models used for modeling the main processes involved in ADME, mostly as they relate to oral absorption. It is evident that most attempts to predict oral bioavailability for diverse sets of pharmaceuticals have failed; therefore many researchers have turned their efforts to investigate the possibility of predicting the various components playing a role in this process. Absorption is definitely one of the controlling components and a number of models of various complexity have been proposed in the literature to predict its value (Agatonovic-Kustrin et al., 2001, Balimane et al., 2000, Clark, 1999, Clark, 2001, Egan et al., 2000, Knutson et al., 2000, Norinder et al., 1999, Palm et al., 1996, Palm et al., 1997, Stenberg et al., 2000, Stenberg et al., 2001, Yu and Amidon, 1999).

One of the simplest and most widely used computational approaches to estimate permeation or absorption was developed by Lipinski, based on an evaluation of the value of certain physical chemical properties of successful drugs. Based on this evaluation, Lipinski proposed the ‘Rule of Five’ (Lipinski et al., 1997) which states that poor absorption or permeation are more likely to be observed when: (1) the molecular weight is greater than 500, (2) the log P is larger than 5, (3) there are more than five H-bond donors, expressed as the sum of all OHs and NHs, and (4) there are more than 10 H-bond acceptors, expressed as the sum of all Os and Ns. However, there are some glaring exceptions which cannot be explained by these rules. The most notable ones are the permeation of orally active therapeutic classes such as antibiotics, antifungals, vitamins and cardiac glycosides. Improvements over the ‘Rule of Five’ by the use of polar surface area (PSA), a parameter closely related to the capacity of a compound to participate in hydrogen bonds, have been reported recently (Clark, 1999).

The first approach to the quantitative prediction of human intestinal absorption for a large set of drugs was obtained from a genetic algorithm combined with a neural network fitness evaluator. The authors, Wessel et al. (1998), elaborated and tested their QSPR model using 86 drug and drug-like compounds, which were encoded with calculated molecular structure descriptors. The calculated %HIA (human intestinal absorption) model had a root-mean-square error (rms) of 9.4% HIA units for the training set, 19.7% HIA units for the cross validation set, and 16.0% HIA units for an external prediction set. These results show that when coupled with a genetic algorithm, neural network is a relatively good alternative approach to the prediction of human intestinal absorption.

In another study, Zhao et al. (2001) put together a solvation equation to model the human intestinal absorption data for 169 drugs. The model containing five ‘Abraham descriptors’ was used to evaluate the human intestinal absorption of 241 drugs, the largest set of data ever used in modeling human oral absorption. They developed several training sets based on different numbers of chemicals. One of these models used 38 drugs, had an r2 of 0.83 and a root mean square error (RMSE) of 14. This model was validated by testing a data set of 131 compounds with good statistical results, RMSE also being 14 units. Another model was built using a set of 169 drugs but with less desirable statistical parameters, r2 being only 0.74.

Thus, as mentioned, there are indeed a few models capable of giving reliable estimates of human intestinal absorption. However, they either predict drug absorption qualitatively or are based on a limited set of compounds. In this paper we discuss a different approach to the evaluation of oral drug absorption, which may help not only with the prediction of the absorption potential of new drug candidates but also in the design of drugs with improved absorption characteristics.

Section snippets

Human intestinal absorption database

Data for our human intestinal absorption model was compiled from various literature sources (Chiou et al., 2000, Irvine et al., 1999, Palm et al., 1997, Raevsky et al., 2000, Stenberg et al., 2001, Wessel et al., 1998, Yu and Amidon, 1999, Zhao et al., 2001). A large amount of data was obtained via the Internet from Gold Standard Multimedia (Reents et al., 2001), and also from a 30-day trial version of the USP DI desktop series (copyright 2000, MICROMEDEX). When the values were given as a

Results and discussion

Due to relatively few reports of drugs with low intestinal absorption in the literature, our learning set is significantly biased towards high activity values. This is a known problem and has been reported by other authors (Wessel et al., 1998, Zhao et al., 2001) who developed models for human intestinal prediction. The mean experimental human intestinal absorption data for drugs used to build the model was 75.97% while the one for compounds present in the test set was 76.43%, which shows that

Conclusions

A QSAR model for the rapid computation of human intestinal absorption was developed. The model includes compounds that are transported by different mechanisms through intestinal membranes. One might argue that a method to calculate absorption for compounds subject to different kinds of transport mechanisms should not be considered.

However, the results we present in this paper represent a confirmation of the validity of previously developed models that were based on human intestinal absorption

References (36)

  • Y.H Zhao et al.

    Evaluation of human intestinal absorption data and subsequent derivation of a quantitative structure–activity relationship (QSAR) with the Abraham descriptors

    J. Pharm. Sci.

    (2001)
  • W.L Chiou et al.

    Evaluation of using dog as an animal model to study the fraction of oral dose absorbed of 43 drugs in humans

    Pharm. Res.

    (2000)
  • D.E Clark

    Prediction of intestinal absorption and blood–brain barrier penetration by computational methods

    Comb. Chem. High Throughput Screen.

    (2001)
  • D Clemett et al.

    Azimilide

    Drugs

    (2000)
  • W.J Egan et al.

    Prediction of drug absorption using multivariate statistics

    J. Med. Chem.

    (2000)
  • D.P Figgitt et al.

    Levosimendan

    Drugs

    (2001)
  • G Klopman

    Artificial intelligence approach to structure-activity studies

    J. Am. Chem. Soc.

    (1984)
  • G Klopman

    multicase: a hierarchical computer automated structure evaluation program

    Quant. Struct.–Act. Relatsh.

    (1992)
  • Cited by (129)

    View all citing articles on Scopus
    View full text