Reviews
Machine learning approaches for predicting compounds that interact with therapeutic and ADMET related proteins

https://doi.org/10.1002/jps.20985Get rights and content

ABSTRACT

Computational methods for predicting compounds of specific pharmacodynamic and ADMET (absorption, distribution, metabolism, excretion and toxicity) property are useful for facilitating drug discovery and evaluation. Recently, machine learning methods such as neural networks and support vector machines have been explored for predicting inhibitors, antagonists, blockers, agonists, activators and substrates of proteins related to specific therapeutic and ADMET property. These methods are particularly useful for compounds of diverse structures to complement QSAR methods, and for cases of unavailable receptor 3D structure to complement structure‐based methods. A number of studies have demonstrated the potential of these methods for predicting such compounds as substrates of P‐glycoprotein and cytochrome P450 CYP isoenzymes, inhibitors of protein kinases and CYP isoenzymes, and agonists of serotonin receptor and estrogen receptor. This article is intended to review the strategies, current progresses and underlying difficulties in using machine learning methods for predicting these protein binders and as potential virtual screening tools. Algorithms for proper representation of the structural and physicochemical properties of compounds are also evaluated. © 2007 Wiley‐Liss, Inc. and the American Pharmacists Association J Pharm Sci 96: 2838–2860, 2007

Section snippets

INTRODUCTION

Modern drug discovery efforts have primarily been focused on the search and optimization of agents interacting with specific therapeutic target, possessing desirable ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties, and exhibiting insignificant adverse drug reactions.1., 2., 3. Methods for predicting these pharmacodynamic and ADMET properties, particularly in early discovery stages, are highly useful for facilitating drug development and drug safety evaluation.1,

DRUG DISCOVERY AND PREDICTION OF COMPOUNDS THAT INTERACT WITH THERAPEUTIC AND ADMET RELATED PROTEINS

Most drugs exert their therapeutic actions by inhibiting, antagonizing, blocking, agonizing, or activating specific therapeutic target protein.6 For instance, many antidepressant drugs target proteins that modulate neurotransmission particularly that of monoamines, which include 5HT reuptake inhibitors, adenosine receptor A agonists, alpha 2 blockers, CRF antagonists, dopamine D antagonists, dopamine reuptake inhibitors, HT agonists, HT antagonists, MAO inhibitors, and norepinephrine reuptake

MOLECULAR DESCRIPTORS FOR REPRESENTING COMPOUNDS

Molecular descriptors are used for representing structural and physicochemical properties of compounds from their 1D, 2D or 3D structure. The most popularly used computer programs for deriving molecular descriptors are DRAGON,67 Molconn‐Z,68 JOELib,69 and Xue descriptor set.25 Web‐servers such as MODEL (http://jing.cz3.nus.edu.sg/cgi‐bin/model/model.cgi) have also emerged for facilitating the computation of molecular descriptors. Over 3000 molecular descriptors can be derived from these

COMMONLY USED MACHINE LEARNING METHODS

Several machine learning methods have been widely used for the classification of pharmaceutical relevance. These include logistic regression (LR), linear discriminant analysis (LDA), k nearest neighbor (kNN), binary kernel discrimination (BKD), decision tree (DT), artificial neural network (ANN), probabilistic neural network (PNN), and support vector machine (SVM). Websites for the freely downloadable codes of some methods are given in Table 1.

METHODS FOR TRAINING, TESTING AND ESTIMATING GENERALIZATION CAPABILITIES OF MACHINE LEARNING METHODS

Several validation methods have been used for training, testing, and estimating generalization errors of a ML model based on a “re‐sampling” strategy.94, 95 The commonly used validation methods include N‐fold cross‐validation, leave one out, leave‐v‐out, jack‐knifing, and bootstrapping. In N‐fold cross‐validation, compounds are randomly divided into N subsets of approximately equal size. N − 1 subsets are used as a training set for developing a ML model, and the remaining one is used as a

SELECTION OF MOLECULAR DESCRIPTORS BY FEATURE SELECTION METHODS

Not all of the available molecular descriptors are needed for representing features of a particular class of compounds. Descriptors most appropriate for representing compounds of a particular property can be selected either by intuition as those used in QSAR and QSPR studies11, 13, 14 or by using feature selection methods. The commonly used feature selection methods include genetic algorithm‐based approach,97 recursive feature eliminations (RFE),98 and simulated annealing‐based approach.99 Some

PERFORMANCE MEASUREMENT

As in the case of all discriminative methods,107 the performance of ML methods can be evaluated by the quantity of true positives TP, true negatives TN, false positives FP (negatives but misclassified as positives), false negatives FN (positives but misclassified as negatives). Here positives refer to compounds having a particular pharmaceutical activity such as inhibitors, agonists or substrates of a protein and negatives refer to those compounds without the pharmaceutical property such as

Prediction of Inhibitors, Antagonists, Blockers, Agonists, Activators, and Substrates of Proteins Related to Specific Therapeutic and ADMET Property

Table 2 summarizes the reported performance in using ML methods for predicting inhibitors, antagonists, blockers, agonists, activators, and substrates of pharmaceutical relevance. The number of compounds in many of the studies listed in Table 2 is in the range of hundreds or even thousands of compounds, which is significantly higher than the tens of compounds typically used in QSAR and QSPR studies108 and closer to those used for developing structure‐based21, 22 and ligand‐based23, 24 VS

UNDERLYING DIFFICULTIES IN THE APPLICATION OF MACHINE LEARNING METHODS

The performance of ML methods critically depends on the diversity of compounds in a training dataset and the appropriate description of the compounds. The datasets used in the most of the ML models described in Table 2, Table 3, Table 4 are not expected to be fully representative of all of the compounds interacting with and those not interacting with a particular therapeutic or ADMET related protein. This is particularly true for compounds not interacting with a therapeutic or ADMET related

CONCLUSIONS AND PERSPECTIVES

ML methods consistently show promising capability for predicting compounds of diverse ranges of structures and of a wide variety of protein binding activities of pharmaceutical relevance. Regression‐based ML methods can be used for quantitative prediction of the activity levels if the activity data are available for a sufficient number of compounds with specific binding activity. Regression methods have the capacity for estimating the contribution of specific structural and physicochemical

REFERENCES (128)

  • S. Ribeiro et al.

    The clinical potential of chemokine receptor antagonists

    Pharmacol Ther

    (2005)
  • D. Fabbro et al.

    Protein kinases as targets for anticancer agents: From inhibitors to useful drugs

    Pharmacol Ther

    (2002)
  • W.K. Oh

    The evolving role of estrogen therapy in prostate cancer

    Clin Prostate Cancer

    (2002)
  • A.P. Li

    Screening for human ADME/Tox drug properties in drug discovery

    Drug Discov Today

    (2001)
  • A.P. Li et al.

    Substrates of human hepatic cytochrome P450 3 A4

    Toxicology

    (1995)
  • D.A. Smith et al.

    Properties of cytochrome P450 isoenzymes and their substrates. Part 1. Active site characteristics

    Drug Discov Today

    (1997)
  • M.J. de Groot et al.

    Pharmacophore modeling of cytochromes P450

    Adv Drug Deliv Rev

    (2002)
  • R.G. Tirona et al.

    Nuclear receptors and drug disposition gene regulation

    J Pharm Sci

    (2005)
  • S. Ekins

    Predicting undesirable drug interactions with promiscuous proteins in silico

    Drug Discov Today

    (2004)
  • L. Schmitt et al.

    Structure and mechanism of ABC transporters

    Curr Opin Struct Biol

    (2002)
  • M.C. Hemmer et al.

    Deriving the 3D structure of organic molecules from their infrared spectra

    Vibrational Spectrosc

    (1999)
  • M. Randic

    Graph theoretical approach to local and overall aromaticity of benzenoid hydrocarbons

    Tetrahedron

    (1975)
  • D.F. Specht

    Probabilistic neural networks

    Neural Netw

    (1990)
  • C.B. Lucasius et al.

    Understanding and using genetic algorithms. Part 1. Concepts, properties and context

    Chemometr Intell Lab

    (1993)
  • J.M. Sutter et al.

    Comparison of forward selection, backward elimination, and generalized simulated annealing for variable selection

    Microchem J

    (1993)
  • J. Drews

    Drug discovery: A historical perspective

    Science

    (2000)
  • J. Caldwell et al.

    An introduction to drug disposition: the basic principles of absorption, distribution, metabolism, and excretion

    Toxicol Pathol

    (1995)
  • R.E. White

    High‐throughput screening in drug metabolism and pharmacokinetic support of drug discovery

    Annu Rev Pharmacol Toxicol

    (2000)
  • C.J. Zheng et al.

    Therapeutic targets: Progress of their exploration and investigation of their characteristics

    Pharmacol Rev

    (2006)
  • C.J. Zheng et al.

    Drug ADME‐associated protein database as a resource for facilitating pharmacogenomics research

    Drug Dev Res

    (2004)
  • Z.L. Ji et al.

    Drug adverse reaction target database (DART): Proteins related to adverse drug reactions

    Drug Saf

    (2003)
  • A.R. Katritzky et al.

    QSPR as a means of predicting and understanding chemical and physical properties in terms of structure

    Pure Appl Chem

    (1997)
  • H. van de Waterbeemd et al.

    ADMET in silico modelling: Towards prediction paradise?

    Nat Rev Drug Discov

    (2003)
  • M.W.B. Trotter et al.

    Support vector machines for ADME property classification

    QSAR Combinatorial Sci

    (2003)
  • X. Yao et al.

    QSAR and classification study of 1,4‐dihydropyridine calcium channel antagonists based on least squares support vector machines

    Mol Pharm

    (2005)
  • C.W. Yap et al.

    Quantitative structure–pharmacokinetic relationships for drug distribution properties by using general regression neural network

    J Pharm Sci

    (2004)
  • G. Harper et al.

    Prediction of biological activity for high‐throughput screening using binary kernel discrimination

    J Chem Inf Comput Sci

    (2001)
  • R.N. Jorissen et al.

    Virtual screening of molecular databases using a support vector machine

    J Chem Inf Model

    (2005)
  • M. Glick et al.

    Enrichment of high‐throughput screening data with increasing levels of noise using support vector machines, recursive partitioning, and laplacian‐modified naive bayesian classifiers

    J Chem Inf Model

    (2006)
  • B.K. Shoichet

    Virtual screening of chemical libraries

    Nature

    (2004)
  • Y. Xue et al.

    Effect of molecular descriptor feature selection in support vector machine classification of pharmacokinetic and toxicological properties of chemical agents

    J Chem Inf Comput Sci

    (2004)
  • H. Li et al.

    Effect of selection of molecular descriptors on the prediction of blood–brain barrier penetrating and nonpenetrating agents by statistical learning methods

    J Chem Inf Model

    (2005)
  • C.A. Bergstrom et al.

    Molecular descriptors influencing melting point and their role in classification of solid drugs

    J Chem Inf Comput Sci

    (2003)
  • J.J. Perez

    Managing molecular diversity

    Chem Soc Rev

    (2005)
  • P. Willett et al.

    Chemical similarity searching

    J Chem Inf Comput Sci

    (1998)
  • T. Potter et al.

    Random or rational design? Evaluation of diverse compound subsets from chemical structure databases

    J Med Chem

    (1998)
  • P. Pacher et al.

    Trends in the development of new antidepressants. Is there a light at the end of the tunnel?

    Curr Med Chem

    (2004)
  • D.O. Stichtenoth et al.

    The second generation of COX‐2 inhibitors: What advantages do the newest offer?

    Drugs

    (2003)
  • M.C. Hochberg

    COX‐2 selective inhibitors in the treatment of arthritis: A rheumatologist perspective

    Curr Top Med Chem

    (2005)
  • L.A. Linkins et al.

    Pharmacology and clinical potential of direct thrombin inhibitors

    Curr Pharm Des

    (2005)
  • Cited by (58)

    • The influence of phase II enzymes on in vitro half-life of pirydo[1,2-c]pirymidine derivatives as structural analogues of arylpiperazine

      2020, Microchemical Journal
      Citation Excerpt :

      The main idea was to validate if a common procedure, well described and used by authors before, can be easily modified to include phase II metabolism. Even though many companies describe the metabolic stability assay and state that simple addition of alamethicin and phase II cofactors can enable phase II investigation of drug metabolism, authors found no evidence regarding how this affects in vitro half-life values [26–28]. According to the chemical equilibrium, when products are constantly “consumed”, the reaction can occur faster.

    • In silico approaches and tools for the prediction of drug metabolism and fate: A review

      2019, Computers in Biology and Medicine
      Citation Excerpt :

      Currently, machine learning is widely used in the field of computer-aided drug discovery, which allows for predicting the interaction between a ligand and a target protein, and hence facilitates the development of new drugs [52–54]. It also seeks to predict the ADMET properties of drugs, thus ultimately facilitating the development of safe and promising agents [55–57]. Drug metabolism is broken down into several phases, each with numerous enzymes that play a role in metabolizing the drug; therefore, a large number of machine learning models have been built to classify a drug's fate based on whether or not the drug will be metabolized by certain enzymes [58].

    • In silico methods for predicting drug-drug interactions with cytochrome P-450s, transporters and beyond

      2015, Advanced Drug Delivery Reviews
      Citation Excerpt :

      Such models could use continuous or classification data. Recently, MLM such as support vector machines (SVMs) has been explored for predicting DDI [51]. P450-mediated DDI is widely known as responsible for a large number of potential drug failures during the preclinical process.

    View all citing articles on Scopus
    View full text