Drug Design 2 - uni-tuebingen.deabi.inf.uni-tuebingen.de/Teaching/.../DD2/Slides/DD2_WS09_12_ADMET.pdf · Drug Design 2 Oliver Kohlbacher Winter 2009/2010 12. ADMET . Overview •

Abt. Simulation biologischer Systeme WSI/ZBIT, Eberhard-Karls-Universität Tübingen

Drug Design 2

Oliver Kohlbacher Winter 2009/2010

12. ADMET

Overview

•  ADME –  Bioavailability

–  Metabolization, Elimination

•  Toxicity –  Effect and side effects

–  Mechanisms of toxicity

–  Models •  Animal models

•  In vitro models

•  Theoretical models and predictions

Failure in Late Development

•  90% of all drug candidates

fail between discovery and

introduction to the market

•  The late development phases

are the most expensive

phases

•  In more than 60% of the

cases, poor pharmacokinetic

(PK) or toxicological properties are the cause

Prentis et al., Br. J. Clin. Pharmacol. 1988, 25, 387-396.

6%

22%

41%

31%

Reason for Failure

Market Toxicity PK Efficacy

Absorption and Elimination

Central Compartment Biliary Renal

Gehirn

Blood-Brain Barrier

Enteral

Absorption

Elimination

Membranes of the GI Tract

Overview of the Different Areas

Application

Dissolution

Absorption

Distribution

Place of Action (Receptors)

Pharmacological Effect

Clinical Effect Toxic Effect

Storage

Biotransformation

Excretion

After: Mut, p. 5

Pharmaceutical Phase

Pharmacokinetic Phase

Pharmacodynamic Phase

Bioavailability

•  Drug has to reach the place of action and achieve a sufficient concentration there for the duration of action

•  This bioavailability is a key criterion for the effectiveness of a drug

•  Also summarized as ADME (Absorption, Distribution, Metabolism, Excretion) and ADMET (ADME-Tox)

–  Pharmaceutical basics: see Lecture Drug Design 1

–  Absorption and Distribution are pharmacokinetic properties

–  Metabolization is much harder to predict

–  It is not only relevant for the prediction of elimination, but also for toxicity (toxic metabolites!)

First Pass Effect

Gastro-intestinal lumen

Gastro-intestinal wall

Liver Blood vessel

Biotransformation Elimination, Biotransformation

•  Absorbed substances have to pass through the gastro-intestinal wall then then through the liver (portal vein)

•  First pass effect: metabolization in the liver before the compound reaches systemic circulation, reduces bioavailability drastically

Biotransformation

•  Oxidation •  Reduction •  Hydrolysis •  Decarboxylation •  Methylation •  Acetylation

•  Conjugation with –  Activated glucuronic

acid –  Sulfuric acid –  Glycine

•  ...

•  Many enzymes catalyze the transformation of substrate families that also include numerous drugs

•  Particularly active in this regard are liver enzymes, which also happen to have a broad substrate specificity

•  Frequent biotransformations are:

Bioavailability

•  Experimental determination of ADME

parameters is time-consuming and costly

•  Computational could thus have a large impact

•  Bioavailability is not governed by a single

property, it is the sum of all ADME processes

•  Modeling it is thus very difficult and QSPR

models still have limited reliability in this

area

Models in Use System Models Pros Cons

In silico (Q)SAR, (Q)SPR

High throughput, cheap, easy to use

Require high-quality exp. data, not all biological processes modeled

In vitro Artificial membranes, cell-based assays (Caco2, MDCK)

Medium to high throughput, includes active and passive transport mechanisms

Many phenomena are strongly model-dependent, no active transport (membrane-based), analytically difficult

In situ Rat intestinal perfusion

Very close to in vivo, includes all key mechanisms except for systemic effects

Labour-intensive, differences between species In vivo Rat portal vein

studies As in situ but also includes presystemic metabolism

Pelkonen et al.; Eur J Clin Pharmacol 57:621-629, 2001

Experimental Models in vitro

in situ

in vivo

http://www.transonic.com http://www.uv.es/~mbermejo/projects.htm

http://www.millipore.com

QSPR Model

•  Turner et al. published a purely computational study for bioavailability in 2003

•  169 with literature data on bioavailability

•  10 cpds. Randomly selected as a test set

•  Training on the remaining 159

•  Model based on eight different descriptors

•  Selected from 94 descriptors by stepwise multiple linear regression

•  Prediction: R = 0.72 Turner et al.; Anal Chim Acta 485;89-102, 2003

QSPR Model

• Descriptors used: –  Electron affinity (H-bonds) –  Number of aromatic rings –  Energy of the highest occupied molecular

orbital (HOMO energy) –  Partition coefficient octanol/water (log P) –  Molar volume –  Ratio of hydrophilic/lipophilic groups –  Solubility in water –  Contribution of H-bonds to solubility

Turner et al., Anal Chim Acta (2003), 485, 89-102

Prediction of Bioavailability

Turner et al., Anal Chim Acta (2003), 485, 89-102

Toxicity and Side Effects

•  Are there drugs without side effects? • W. Kuschinski:

“If it is claimed that a substance has no side effects, then it is to be assumed that it has no desired effect either.“

•  Required: Estimate of toxicity

Toxicity and Side Effects

•  Paracelsus: „The dose makes the poison“

•  No clear distinction between ‘medicine’ and

‘poison’

•  Usually, there is not a single ‘cause’ for a

toxic effect

•  Many mechanisms are involved

•  Toxic effect often also occurs through

metabolization of the substance

Dose-Response Relationships

•  „The dose makes the poison“ –  Typically, an increase in dose

increases the effect –  Above a certain dose, additional

toxic effects may be observed –  Strength/duration of effect depends

on many factors, e.g., genotype, age, body mass, …

–  Difficult to quantify strength of the response

) Dose-response relationships

measured for collectives –  ED50: median effective dose –  LD50: median lethal dose

Katzung, Basisc and Clinical Pharmacology, p. 30

Therapeutic Index

Mut, S. 81

ED50 LD50

•  Therapeutic index or therapeutic ratio is the ratio between the concentration causing a toxic effect and the concentration causing a therapeutic effect

•  It is a measure of a drug‘s safety

Human Toxicity

•  Definition of acute LD50 in human not helpful – it is generally not experimentally accessible

•  A lethal dose has to be avoided at all costs

•  For a safe drug, we need to achieve less than 1 death per million, i.e. LD0,000001

•  Apart form acute toxicity, long-term effects are of great importance

–  Mutagenicity

–  Carcinogenicity

–  ...

Poison/Drug •  Poisons obviously have a biological activity

•  We can thus apply the same principles as for pharmaceutical activity

–  Toxicokinetics

–  Toxicodynamics

•  In principle, all of the methods described to model biological activity are applicable as well

•  Problem:

–  Toxicity is often not a single, well defined process (in contrast to binding, activation, ...)

–  What is the mechanism? Where to start?

Types of Toxicity

• Acute toxicity –  Exposition to a single dose or multiple

doses in a short space of time –  Symptoms occur immediately or briefly

after exposition

• Chronic toxicity –  Prolonged exposition to low-level doses –  Slow accumulation of the poison to toxic

concentrations –  Often due to lack of excretion/elimination

Classification of Poisons

• Target organ (liver, kidneys, ...)

• Application (pesticide, solvent, food

supplement, ...)

•  Source (animal or plant poisons,

synthetic, ...)

•  Effect (cancer, mutagenesis, liver

damage, kidney failure, ...)

Exposition to Poisons

•  Environmental exposition is a common source

•  Exposition to toxic substances is controlled by legislation –  Short-term exposure limit (STEL) [in Germany:

MAK (maximale Arbeitsplatzkonzentration)]

•  However:

Carcinogenic substances are never harmless!

Even smallest amounts can cause genetic alterations.

Molecular Toxicology

•  Studies the interaction of a poison with a

biological object

•  Consider the effect on a molecular level

•  Becoming increasingly important as the

mechanisms of toxic action of compounds are

becoming known

•  Obviously of the utmost importance in the

drug design process

What causes toxicity?

Many different mechanisms involved: •  Biotransformation (metabolization) •  Interactions between several

substances which are not toxic by themselves (at the individual doses)

•  Inhibition/inactivation/denaturation of proteins (enzymes, receptors, ...)

•  Saturation of metabolism •  Activation/blocking of receptors •  ...

Paracetamol •  Very common analgesic and antipyretic

(Acetaminophen)

•  Normal dose: 500-1000 mg

•  Usually well tolerated

•  But: high doses can be toxic!

•  Doses of 10 g/day or more lead to severe live cell necrosis (often lethal)

•  The effect is caused by a toxic metabolite of paracetamol, N-acetyl-p-benzoquinoneimine

The dose makes the poison

Paracetamol (Acetaminophen)

Paracetamol

•  Damage to liver cells is caused by highly reactive metabolites

•  Formed by cytochromes, their reaction with proteins in the liver causes the toxic effect

•  Low doses: metabolites are captured and detoxified by glutathione by forming harmless conjugates

•  Toxic dose: –  Glutathione storage exhausted

–  Metabolites cannot be detoxified

Glutathione

Paracetamol

•  Stimulation of chytochrome p450 lowers the glutathione level

•  Some therapeutically active compounds activate p450 (strongly)

•  Consequence: even very low (normally non-toxic) doses of paracetamol become toxic

•  Toxic doses well below 6 g •  Problematic for patients with pre-existing

damage to the liver (alcohol abuse!) ) Interaction with other substances can cause toxicity!

Drug Interactions

•  Example: administration of phenytoin (an

antiepileptic) and salicylic acid at the same time

results in abnormally high plasma levels

•  Problem: both drugs are eliminated through the same

enzyme

•  Consequence: saturation of the enzyme reduces

elimination rates

•  This can lead to toxic effects, as the effective plasma

concentration is much higher than anticipated

Drug interaction by saturation of an enzyme!

Benzene/Toluene

•  Benzene has a high acute toxicity and is also carcinogenic – in contrast to structurally very similar toluene

•  Structures differ only by a single methyl group

•  Benzene is easily absorbed, also across the skin

•  It is also easily excreted again, mostly through the lung

•  About half of the absorbed amount is typically metabolized

•  Toxic effect is most likely due to this biotransformation

Benzene

Toluene

Benzene

Acute toxicity

•  More than 0.5 ml/kg causes

–  Intoxication

–  Headaches, dizziness

•  Higher doses:

–  Convulsions

–  Unconsciousness

–  Cardiac arrhythmia

–  Eventually death by central respiratory paralysis

Benzene

Chronic toxicity or massive single doses:

•  Hemotoxicity –  Inhibition of erythropoiesis, leukopoiesis and

thrombopoiesis

–  No therapy known

•  Carcinogenicity –  Leukemia

–  Causes irreversible chromosomal aberrations in lymphocytes and bone marrow cells

–  Benzene is one of the most important environmental poisons

Metabolization of Benzene

•  Initial step: enzymatic epoxidation

•  Epoxides are highly reactive

•  Can react with hydrogen atoms of biological

macromolecules

•  Carcinogenic and mutagenic effects is

(probably) caused by reactions with nucleic

acids

•  Toluene is metabolized differently (starting

with the methyl group) and is thus much less

toxic

Mono- oxygenase

Carcinogens

•  In animal models, symmetric dialkylnitrosamines cause

–  Liver tumors after chronic exposure to low doses

–  Kidney tumors after exposure to a single high dose

•  Small amounts of nitrosamines are very common (in particular in alcoholic beverages, certain meat products)

•  Also formed endogenously (production of nitrite from saliva and gastric juice)

Dimethylnitrosamine

Carcinogens •  Polycyclic aromatic hydrocarbons

(PAHs) are often carcinogenic

•  Effect is again caused by

metabolites

•  Some of these metabolites show

acute gene toxicity

•  They have also been shown to

react with DNA in vivo

•  Bay region is important for

metabolic activation

Benzo(a)pyren

1,2-5,6 Dibenzanthrazen

Metabolization

Metabolization of benzo(a)pyrene to a diol epoxide, which then reacts with the exocyclic amino group of guanine

Testing Prevents Disasters •  No or insufficient testing of novel pharmaceutical

compounds led to several major disasters –  Brain damage and death in small children due to

sulfonamides (late 1930s) –  More than 100 deaths through diethylene glycol used as a

solvent for sulfanilamide (this incident led to the foundation of the Food and Drug Administration [FDA] in the USA)

–  Severe birth defects after the use of thalidomide during pregnancy in about 10,000 children worldwide

•  High standards for drug safety have drastically reduced these incidents

•  Nevertheless, drugs are taken off the market again because long-term (side) effects have not been recognized early on

Models: Animal Models

•  How to test for human toxicity early on?

•  Difficult: there is rather little toxicological data

available for humans (systematic toxicological testing

with human subjects is not considered acceptable!)

•  The vast majority of reliable data thus stems from

animal experiments

•  As we have seen before, these data are often hard to

transfer to humans

•  But: better than nothing!

Animal Models

•  Animal experiments are...

–  Expensive

–  Time-consuming

–  Raise ethical issues

–  Required by law

•  Reduction of animal use by

–  In vitro models

–  Computational models

http://abclabs.com

Animal Use

•  Strong growth of animal use between 1945 and 1968

•  Stagnation until the middle of the 70s, then steady decrease (1978-1988 decrease by 60% in West Germany)

•  Reduction quite remarkable: more substances tested than ever!

•  Key reason: in vitro tests (Ames test)

•  Theoretical models still play a very minor role

Animal Use 1991-1995 Species 1991 1992 1993 1994

Mouse 1.223.741 1.064.883 973.106 868.312

Rat 611.530 558.516 508.769 459.781

... ... ... ... ...

Total 2.402.710 2.082.588 1.924.221 1.758.500

Quelle: Bundesministerium für Verbraucherschutz, Information Nr. 44 v. 30. Oktober 1995 http://www.bmelv.de/cae/servlet/contentblob/765788/publicationFile/43424/2008-TierversuchszahlenGesamt.pdf

•  About half of them are related to medical research

•  Most of them are rodents (mice, rats)

•  Over the last decade the number of animals used in animal experiments has been increasing steadily (2008: 2.6 mio. in Germany)

•  Number for mice are always increasing, most other species going down

•  About 171,000 are currently being used per year for toxicological studies

Comparison between Species •  Different species may react very differently to the same drug

•  Example: lysergic acid diethylamide (LSD)

–  Experiment: administer a hallucinogenic, but subtoxic, dose to a male Asiatic elephant

–  Estimate: dose of 0.3 g

(about 0.06 mg/kg)

–  Result: Death.

) Toxic dose for an

elephant about 1000

lower than for a mouse!

West LJ, Pierce CM, Thomas WD. Lysergic Acid Diethylamide: Its Effects on a Male Asiatic Elephant. Science. 1962 Dec 7;138(3545):1100-1103.

Transferability of the Data

Toxicity of LSD

Species LD50 [mg/kg]

Mouse 50-60

Rat 16.5

Rabbit 0.3

Elephant << 0.06

Human >> 0.003

Transferability of Data

•  Example: 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD, “Seveso poison”)

•  Even in the closely related species hamster and guinea pig toxicity differs by three orders of magnitude

•  Extrapolation is thus very, very dangerous!

•  Even data from close relatives can be misleading: apes are rather insensitive to TCDD

Models: in vitro Tests

•  Example: Ames test identifies cpds with mutagenic (and carcinogenic) potential

•  Carefully engineered strain of Salmonella typhimurium

•  Lacks the ability to synthesize His •  Incubated together with the compound and a few

other things (e.g., liver extract to check for mutagenicity of possible metabolites)

•  Mutagenic agent can cause backmutations that can grow into larger colonies

•  Colony count is a measure for mutagenic potential Ames, B., F. Lee, and W. Durston; Proc. Natl. Acad. Sci. USA 70:782-786, 1973

Problems Predicting Toxicity

•  Wide range of biochemical processes involved

•  Very similar structures have very different toxicological properties (structure-toxicity landscape is very rough)

•  Different mechanisms can result in the same toxicological outcome

•  Very different structure-activity relationships between different classes of compounds

•  Often caused by metabolites, so not a property directly related to a compound‘s structure!

Theoretical Models

Which approaches are there?

•  Knowledge-based (“expert systems”)

–  Store expert knowledge in as individual rules

–  Applying these rules to a given structure results in a classification

•  Statistical models ((Q)SAR)

–  Automated statistical analysis of large-scale data sets using statistical methods

–  No experts required, but strongly dependent on data quality, inclusion of all relevant processes

Top-down/Bottom-up

Richard, Toxicol Lett 102-103:611-616, 1998

What is being modeled? •  There is not an single toxicity model, but

numerous different toxicological properties are being modeled independently: –  Liver toxicity

–  Kidney toxicity

–  Carcinogenicity

–  Mutagenicity

–  Reproductive toxicity

–  Acute LD50 (rat)

–  Transdermal absorption

–  ...

Approaches Statistical Approaches

•  CASE/MultiCASE

Klopman; J. Am. Chem. Soc 106:7315-7321, 1984

Klopman; QSAR 11:176-184, 1992

•  TOPKAT

Enslein et al.; Mutat. Res. 305:47-61, 1992

Expert Systems

•  DEREK

Sanderson, Earnshaw; Human Exp. Toxicol. 10:261-273, 1991

•  ONCOLOGIC

Woo et al.; Toxicol. Lett. 79:219-228, 1995

Prediction Models

Richard, Toxicol Lett 102-103:611-616, 1998

Example: DEREK

•  DEREK is a knowledge-based system (Deductive Estimation of Risk from Existing Knowledge)

•  Based on a program for organic synthesis planning (LHASA)

–  Rather old(1980 bzw. 1985)

–  Developed for VAX

–  600 modules in Fortran, C and Macro

–  Rule-description language CHMTRN

•  Extends LHASAs rule language for elements required in toxicology (DERTRN)

•  Initially about 50 different rules

„...based on a combination of over 30 years experience of toxicological work...“

DEREK •  Simple rules of the structure IF structural chemical property THEN specific outcome possible

•  Results are purely qualitative (!) •  In addition, it contains the FDA‘s rule

set for carcinogenicity • Mainly used to select compounds in a

campaign, remove those with obvious problems

•  Also used as an indicator where additional experiments are required

DEREK Rules

Four sections

1.  Descriptions

2.  Usage information

3.  Structural pattern

triggering the rule

4.  DERTRN query to refine

the structural pattern

further

DEREK: Pros/Cons

+ Development of rules is overseen by the users and transparent to them – there is always an explanation for a rule

+ New rules can be integrated rather simply

+ Simple user interface

-  Rule syntax rather limited, no 3D definitions

-  Rules are rather coarse and capture only a few key metabolic mechanisms

-  Inclusion of toxicological databases and the knowledge therein might improve predictions

Example: (Multi)CASE

•  CASE - Computer Automated Structure Evaluation

•  Has been extended into Multi-CASE

•  Relies on statistical analysis of a training data set combining compound structure and their biological activity

•  Training data set must contain a broad range of different structures and toxicological endpoints

CASE: Training Data

•  Training data set: relates structures to biological/toxicological activity

•  Activity is given in ‘CASE units’ 10-19: inactive 20-29: weak activity 30-99: active

•  Structures is given as SMILES, KLN, or as a MOL file

•  Analysis is based on heavy atoms alone

CASE: Training Data •  Training data set has to be examined thoroughly to

ensure even coverage of chemspace

–  Check for overrepresented structures

–  Identify important missing structures/mechanisms

•  As many data points as possible should be included

•  Data of very similar cpds. with identical mechanisms can be pooled

•  Problem: each change in the data set leads to a different predictive model

CASE: Model Construction

•  Decompose all structures into fragments of

size 2 – 10

•  Classify fragments as

–  Biophores/toxicophores (statistically active)

–  Biophobes (statistically inactive)

•  Compute physicochemical descriptors and 2D

descriptors for a QSAR analysis

CASE: Classification

• Many fragments are not by themselves

determining factors for toxicological

activity

•  Assumption: binomial distribution of all

fragments within a class

•  Statistically significant deviation from

this distribution fragment is relevant

for activity

CASE: Prediction

•  Two steps: –  Activity prediction –  Estimation of toxicity

•  Input structure is decomposed into fragments

•  Comparison of fragments to biophores/biophobes

•  Prediction of activity likelihood based on these matches

CASE: Prediction

•  Toxicity estimation uses QSAR •  Model based on multivariate analysis •  Forward selection of descriptors •  Initial descriptors

–  Biophores –  Biophobes –  Predicted log P

•  Descriptors are added until the model does not improve

•  All standard caveats discussed earlier in the lecture of course apply!

Multi-CASE

•  Aims at reducing problems caused by highly correlated descriptors

•  Solution: several stepwise CASE predictions –  Predict strongest biophore

–  Remove molecules containing this biophore from the training data

–  Repeat until the training data set is empty or no significant improvement can be reached

Multi-CASE

•  Distinguishes between activity and modulation of activity –  Split training data set into different classes, based on

presence or absence of a biophore –  Conduct a QSAR analysis in each of the classes to

determine whether related biophores lead to an increase/decrease in activity

•  Uses a larger number of descriptors than CASE •  Prediction:

–  Search for biophores in the input structure –  For each biophore identified, search for modulating

biophores

Multi-CASE: Pros/Cons

+ Predictive models do not require prior

(expert) knowledge

+ Quantitative prediction

-  Prediction accuracy critically dependent on

the quality (and manual curation) of the

training data set

-  Output often ambiguous expert needed!

Multi-CASE vs. DEREK

Prediction accuracy [%]

DEREK 59

Multi-CASE 49

COMPACT 54

TOPKAT 57

Greene; ADDR 54:417-431, 2002

Prediction of carcinogenicity of 44 compounds

Multi-CASE vs. DEREK

DEREK 4.01 Multi-CASE 3.45

Sensitivity 45% 30%

Specificity 62% 84%

Concordance 60% 79%

Greene; ADDR 54:417-431, 2002

Prediction of Ames test results for 974 cpds.

State of the Art – 1990

•  In 1990 Tennant et al. asked several experts in

toxicology to predict carcinogenicity for 44

compounds

•  Computational methods were applied in parallel

•  After experimental testing, the following results

were obtained in 1993:

–  Best result: expert (80% correct)

–  Theoretical approaches: 45-65% correct

•  Results hardly better than random!

•  Not good enough to replace animal models! Tennant et al.; Mutagenesis 5:3-14, 1990

Ashley, Tennant; Mutagenesis 9:7-15, 1994

State of the Art – 2003

•  Predictive Toxicology Challenge: Competition to assess the quality of modern in silico methods

•  Machine learning

•  Limited to the prediction of carcinogenicity

•  Fourteen teams contributed predictions

•  111 models

•  Result: five(!) models performed better than random

Tuivonen et al.; Bioinformatics 19(10):1183-1193, 2003

State of the Art – 2009

•  Valerio et al. (FDA) examined two popular state-of-the-art software packages

–  LMA (Leadscope Model Applier)

QSAR/data mining approach based on structural features

–  MC4PC (a MultiCase descendant)

Rule-based and QSAR approach

•  An external dataset of 43 phytochemicals with known rodent carcinogenicity was used to validate the predictions

Valerio et al., Mol. Nutr. Food Res., 54:1-9 (2010) Yang et al., Toxicol. Mech. Methods, 18:277-295 (2008)

Matthews et al., Toxicol. Mech. Methods, 18:189-206 (2008)

State of the Art – 2009

•  Results: comparable for both programs

•  High specificity, low sensitivity

•  Still not very convincing

•  Combination of both codes into a consensus leads to even worse predictions

Valerio et al., Mol. Nutr. Food Res., 54:1-9 (2010)

MC4PC LMA

Specificity 94% 59%

Sensitivity 47% 50%

False positives 6% 41%

False negatives 53% 50%

Are predictions possible?

•  At the moment predictions are not sufficiently

reliable

•  Even modern statistical learning methods fail to

capture the full complexity of toxicology

•  Without human experts and experimental testing, no

reliable statement on a compound’s toxicity is

possible

•  Nevertheless:

–  Good tool for the expert to guide toxicological studies

–  Can yield important hints for an early selection of

candidates

Summary •  Bioavailability is an essential property for a drug

•  In silico predictions are possible, although difficult using QSAR approaches

•  All drugs have toxic side effects

•  Important is the therapeutic index

•  Prediction is very difficult due to the complexity of toxicological mechanisms

•  in vitro and in silico approaches still cannot replace animal models

•  In silico toxicity predictions: knowledge/rule-based and statistical approaches are currently in use

•  Predictions are still not reliable enough, although they are being used to guide decisions

References Books •  [BKK] Böhm, Klebe, Kubinyi: Wirkstoffdesign, Spektrum 2002 •  Mutschler: Drug actions. Basic Principles and Therapeutic

Aspects, Medpharm Scientific Publishers; Auflage: 6Rev Ed (1994)

•  Klaassen: Casarett and Doull's Toxicology: The Basic Science of Poisons, Mcgraw-Hill Professional; 7th revised ed. (2008)

Papers •  Valerio LG Jr, Arvidson KB, Busta E, Minnier BL, Kruhlak NL,

Benz RD. Testing computational toxicology models with phytochemicals. Mol Nutr Food Res. 2009 (Epub ahead of print), PMID: 20024931