Strategic Testing

CZ3253: Computer Aided Drug designCZ3253: Computer Aided Drug design

Lecture 10: Overview of Drug Testing Methods II: Lecture 10: Overview of Drug Testing Methods II: Test of TOXTest of TOX

Prof. Chen Yu ZongProf. Chen Yu Zong

Tel: 6874-6877Tel: 6874-6877Email: Email: [email protected]@nus.edu.sghttp://xin.cz3.nus.edu.sghttp://xin.cz3.nus.edu.sg

Room 07-24, level 7, SOC1, Room 07-24, level 7, SOC1, National University of SingaporeNational University of Singapore

22

StrategicTesting

SelectValidationChemicals

QSARModelsRules

NN,SVM

StructuralRequirementsFor Pathways

RegulatoryAcceptancy

Criteria

QSAR LibrariesRules Collections

NN, SVM ClassifiersModeling Engine

Estimation of Missing Data

Analogue Identification

Prioritization/Ranking

DistributedDatabases

ChemicalInventories

Identify early critical events

High QualityData Sets

Toxic EffectsPathways

33Library of Toxicological Pathways

Init

iati

ng

Eve

nts

Imp

aire

d R

epro

du

ctio

n/D

evel

op

men

t

Mapping Toxicological Paths to Adverse Outcomes “Estrogen Signaling Pathway”

ER Binding

ER Transactivation

VTG mRNA

Vitellogenin Induction

Sex Steroids

Altered Reproduction/Development

Molecular Cellular Organ Individual

Chemical 3-D

Structure/Properties

Chemical 2-D

Structure

Structure

44

Analysis of Commercial Computational Analysis of Commercial Computational Toxicology SoftwareToxicology Software

• QSAR-BASED– Collects Molecular Fragments and

Descriptors– Calculates Values of Chemical

Descriptors– Compares to Known Compounds– Reports Probability of Being a Member

of a Toxic Class Using Multifactorial Statistical Analysis

– Identifies Structural Liabilities– Unvalidated Structural Relationships

• EXPERT RULE-BASED– Inspects Molecules for Known Structural

Liabilities

– Identifies Structural Liabilities

– Prepares Summary Report of Findings

– Validated Structural Relationships with Known Toxic Mechanisms

– Provides References & Predicted Mechanisms

ADAPT/TOPKAT MultiCASE/LeadScope DEREK

55

Strengths and Weaknesses of Strengths and Weaknesses of Virtual Toxicology Commercial SoftwareVirtual Toxicology Commercial Software

• QSAR-BASED– Provide Relative Dose and Liability

Prediction– Easy to Determine if Compound is Well

Represented in Training Set via Similarity Search

– Can Be Biased to Minimize False Positives and/or False Negatives

– Challenging to Systematically Improve Model: No Linearity

– Difficult to Train General Model: Excellent Predictiveness for Single Event; Problematic for Multiple Events

Good For Specific Models

• EXPERT RULE-BASED– Chemically Intuitive Results

– Good Initial Filter for Known Liabilities: Lacks Specificity

– Only Predicts Presence of Identified Fragments

– Cannot Discriminate within a Structural Sub-Class

– Retrospective in Nature

– Cannot Extrapolate Prediction to New Chemotypes

Good For General Models

ADAPT/TOPKAT MultiCASE/LeadScope DEREK

66

New Toxicology Prediction MethodsNew Toxicology Prediction MethodsNeural Networks– Use of structure descriptors to discriminate between modes of toxic action of

phenols. J Chem Inf Model. 2005 Jan-Feb;45(1):200-8. – Toward an optimal procedure for PC-ANN model building: prediction of the

carcinogenic activity of a large set of drugs. J Chem Inf Model. 2005 Jan-Feb;45(1):190-9.

SVM– Prediction of torsade-causing potential of drugs by support vector machine

approach. Toxicol Sci. 2004 May;79(1):170-7. Epub 2004 Feb 19.– Prediction of Genotoxicity of Chemical Compounds by Statistical Learning

Methods. J Chem Inf Model

Fuzzy Set– Prediction of noninteractive mixture toxicity of organic compounds based on

a fuzzy set method. J Chem Inf Comput Sci. 2004 Sep-Oct;44(5):1763-73.

Ensemble recursive partitioning– In silico models for the prediction of dose-dependent human hepatotoxicity.

J Comput Aided Mol Des. 2003 Dec;17(12):811-23.

77

Predictive GenotoxicityPredictive Genotoxicity

• Goal: Improve current predictive toxicology capabilities for mutagenicity and carcinogenicity through customizing and augmenting current predictive software

• 1. Modeling & Informatics:– Enhancing current predictive software.

• Bias model to minimize false negatives (and indeterminants).

– Provide support to discovery groups to eliminate mutagenic liabilities.– Create a central data repository and populate it with literature data as well as institutional data.– Deliver a predictive mutagenicity package in a format that can be supported as a standard

system.– Allow for novel models to be added as they are developed

• 2. Use:– Prioritization of synthesis & testing candidates.– Identification of substructures responsible for an observed mutagenic liability and suggested

synthetic alternatives.– Regulatory and due diligence support (what will the FDA see?).

88

Requirements for a Requirements for a QualityQuality Filter Filter

• Identify ALL compounds having mutagenic liability– Identify strengths & weaknesses of models– Identify strategy for maintaining & improving the model– User friendly & intuitive– Provide support information for model

• Chemists’ and toxicologists’ needs are not always equivalent– Chemistry:

• Suggest synthetic alternatives; do not limit chemical space• Repository of prior knowledge (both institutional and external)

– Toxicology:• Prioritization of synthesis and in vitro testing candidates• Regulatory and due diligence support; overprediction is acceptable

99

Annotation of Substructural AlertsAnnotation of Substructural Alerts 95 mutagenicity alerts annotated

76 Native DEREK mutagenicity alerts 6 reclassified carcinogenicity alerts (genotoxic mechanism) 13 alerts Implemented by BMS ~300 DEREK Literature References Extracted, Archived and Summarized Probable mechanism(s), including reactive intermediates, described Additional SARs & mechanisms derived using publicly available data (TOXNET, RTECS, NTP) Updated literature archived, integrated and summarized

300+ additional references Lessons learned from QSARs included Validation Statistics included

1010

2002 Validation 2002 Validation Expanded Data SetExpanded Data Set

Class % Sample # CompoundsBMS 23% 416 (65 / 351)

Drugs 29% 534 (107 / 427)

Other 48% 875 (398 / 477)

Ames Pos: 31% 570

Ames Neg: 69% 1255

•~5% of BMS space covered by validation compounds.•~10% of drug space covered by validation compounds.

All Data:

1825 Compounds

Drugs:

534 Compounds

BMS:

416 Compounds

1111

0 100 200 300 400 500 600 700 800

False (+)

False (-)

Indeterminate

Concordance

DEREK (DK) TOPKAT (TPK) MultiCASE (MC)

Parallel DK/MC Parallel DK/TPK Parallel MC/TPK

Parallel DK/MC/TPK Sequential D/MC Sequential MC/TPK

Sequential DK/MC & DK/TPK Sequential DK/MC/TPK

Which program works best?Which program works best?A combination of twoA combination of two

Random

Random

1212

Improving the System:Improving the System:Correction of the 2Correction of the 2oo Amine Alert Amine Alert

• 20/51 (39%) compounds triggering DR005 were predicted positive by an Ames assay (S9)

• Derek Rule 005 Addendum– Exclude Secondary Amides– Exclude Secondary

Sulfonamides

• Modified DR 005 Correctly Predicts 20/35 Compounds (57% concordance).

• Reduced False Positives from 31 to 15.

• Additional Rules and QSARs can be Developed to Improve the Accuracy of this Rule Even Further.

Substructure Name # Pos # Occurrences

R'

N O

R

H

SecondaryAmides

0 10

R'

N

S

O

H

OR

SecondarySulfonamide

0 8

N

R

H

Ar

SecondaryAniline

Derivatives19 30

N

R

H

R'

SecondaryAmines

1 5

1313

Improving the System:Improving the System:Substructures Identified by BMSSubstructures Identified by BMS

Alert NameAmes

PositiveBMS DEREK

PositivePositive

AccuracySubstructure 1 42 77 54.5%

Substructure 2 41 167 24.6%












1414

Predictive ToxicologyPredictive ToxicologyComparing Apples to ApplesComparing Apples to Apples

Secondary and Aromatic Amines: The Data Set: 334 Compounds

Selected for drug-likeness (expanded Lipinski filter) Clustered for diversity Commercially available from Aldrich at over 96% purity

Assayed in the SOS Chromotest assay for genotoxicity Induction of lacZ reporter gene under transcriptional control of SOS DNA

damage repair pathway 90% concordance with the Ames Assay High Reproducibility (± 0.05 fold) 193 compounds considered non-toxic 72 compounds considered weakly toxic 69 compounds considered strongly toxic

1515

Comparing Bad ApplesComparing Bad ApplesMethod % Concordance % Sensitivity % Specificity

ADAPT 72 69 74

TopKat (v5.0) 60 54 63

MultiCASE (A2I) 59 61 57

MultiCASE (SOS) 64 64 64

Leadscope† 74 65 83

DEREK (v5.0) 41 100 0

† Selected Leadscope fingerprints were combined with scaffolds and 8 properties.Logistic PLS method (50 factors) was used after selecting features – Preliminary Data.

1616

Improving Bad ApplesImproving Bad Apples

•You have a positive assessment, now what?

– Correct Molecular Context?• Supporting data?

– Interpolating or Extrapolating?• Is compound within model’s scope?

– Mechanistic Support?• Does the biochemistry make sense?

– Confirmatory Assay• Positive

– Develop with caution• Negative

– Feed data back into model(s)

1717

Exploration of New MethodsExploration of New MethodsStudy Method No. of

CompoundsGT+

Accuracy (%)GT-

Accuracy(%)Q

Overall Accuracy (%)

Snyder RD MCASEDEREKTOPKAT

394394394

48.151.943.4

95.175.188.1

89.673.681.7

Philip D. Mosier k-NN 140 66.7 92.9 85.0

Linnan He Consensus model developed with k-NN, LDA, and PNN classifiers

227 73.8 84.3 81.2

Brian E. Mattioni k-NN 334 69.3 74.1 72.2

BIDD group at NUS

C4.5PNNk-NNSVM

860860860860

55.674.170.477.8

75.080.286.592.7

70.778.982.989.4

1818

Another Example of Toxicity PredictionAnother Example of Toxicity Prediction

• Torsades de Point (TdP): A dangerous side effect of drugs which commonly act at ion channels in the heart to cause arrhythmia

• A common feature of many compounds is activity at the HERG channel (K+)

• Commonly, this is observed as an elongation of the so-called QT interval in an electrocardiogram of the heart (LQT)

1919

Example of SertindoleExample of Sertindole

atypical antipsychotic drug for schizophrenia

- licensed in UK May 1996

– prolonged QTc interval

– cardiac arrhythmias

– by November 1998, MCA/CSM received reports of 36 death and 13 serious but non-fatal arrhythmias

2020

LQTS (long QT Syndrome) LQTS (long QT Syndrome) Data SetData Set

• 122 Compounds, classified into four classes– Class 1: Drugs with Risk of Torsades de Pointes– Class 2: Drugs with possible Risk of Torsades de

Pointes– Class 3: Drugs to be Avoided by Congenital Long QT

Patients– Class 4: Drugs Unlikely to cause Torsades de Pointes

• Using subsets 1 and 4, double cross validation

2121

LQTS Prediction Results LQTS Prediction Results

by by Naïve Bayesian ClassifierNaïve Bayesian Classifier

• Best: ~80% correct predictions• Database not “finalized”• Confusion matrix

Predicted positive

Predicted negative

Compounds from positive set …

30 4

Compounds from negative set …

8 33

2222

Over-Predictions (1)Over-Predictions (1)

Fexofenadine

(“Allegra”, IC50 13m)

Example of Fexofenadine.It’s a modification of terfenadine which has a tertiary butyl group instead of a carboxylic acid. This lowers logD, and therefore it falls below the range of logD necessary for activity.

OH

N

OH

X

NH

N

N

H

H

X=CH3, SeldaneX=COOH Allegra(more hydrophilic)

2323

Over-Predictions (2)Over-Predictions (2)

Example: Sildenafil (Viagra). Positive in Analysis, but negative for Torsades.

Therapeutic ratio is high, 3.5nM at PDE5,But 100m at HERG.

2424

Under-PredictionsUnder-Predictions

Erythromycin A (membrane disrupting)

Loradatine (hydrolysed)

2525

SVM Prediction of TdP SVM Prediction of TdP

Training Testing Independent validation

TdP+ TdP- TdP+ TdP-

TP FN Accuracy %

TN FP Accuracy

%TP FN Accuracy

%TN FP Accuracy

%

64 103 11 0 100.0 29 1 96.7 9 1 90.0 28 1 96.6

Strategic Testing

Documents

j chem inf comput

prediction of dose

bias model

predictive mutagenicity

prediction of genotoxicity

known liabilities

predictive genotoxicitygoal

current predictive software1