Introduction on QSAR Introduction on QSAR and modelling of and modelling of physico-chemical and physico-chemical and biological propertiesbiological properties

Alessandra Roncaglioni – IRFMN

aroncaglioni@marionegri.it

Problems and approaches in computational chemistry

OutlineOutlineHistoryQSAR/QSPR steps

◦(Descriptors)

◦Activity data

◦Modelling approaches

◦Validation (OECD principles)

QSPR (Phys-chem properties)QSAR (Biological activities)Example (DEMETRA)

2

QSAR postulatesQSAR postulatesThe molecular structure is

responsible for all the activities Similar compounds have similar

biological and chemico-physical properties (Meyer 1899)

Hansch analysis (‘70s)Free Wilson approach (‘70s)

H. Kubinyi. From Narcosis to Hyperspace: The History of

QSAR. Quant. Struct.-Act. Relat., 21 (2002) 348-356. 3

Hansch analysisHansch analysisApplied to congeneric series

Log 1/C = a + b + c Es + const.

where

C = effect concentration

= octanol - water partition coefficient

= Hammett substituent constant (electronic)Es= Taft’s substituent constant

Linear free energy-related approach

McFarland principle4

Free-Wilson analysisFree-Wilson analysis

Log 1/C = ai + where C = effect concentrationai= contribution per group

=activity of reference compound

5

The old QSAR paradigmThe old QSAR paradigm

Compounds in the series must be

closely related

Same mode of action

Basics biological activity

Small number of “intuitive”

properties

Linear relation6

The old QSAR paradigmThe old QSAR paradigm

Factors limiting to the old Factors limiting to the old

paradigm:paradigm:

Sw availability

Calculation of molecular properties

Limited COMPUTING POWER

Costs of hw and sw

7

The new QSAR paradigmThe new QSAR paradigm

Heterogeneous compound sets

Mixed modes of action

Complex biological endpoints

Large number of properties

Non linear modelling

8

The new QSAR paradigmThe new QSAR paradigm

Factors enabling new paradigm:Factors enabling new paradigm:

Increased computing power

QM calculations

Thousands of descriptors

Cost drop for hw and sw (freeware)

9

OutlineOutlineHistoryQSAR/QSPR steps

◦(Descriptors)

◦Activity data

◦Modelling approaches

◦Validation (OECD principles)

QSPR (Phys-chem properties)QSAR (Biological activities)Example (DEMETRA)

10

QSAR flowchatQSAR flowchat

11

2D 3D

… … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … …

Descriptos (1, …, m)

D(n,m)

… … … … … … … … … … … … … …

Activity

A A = f (D(n,m))

12

QSAR/QSPR defined by Y QSAR/QSPR defined by Y datadataQuantitative Structure-Property

Relationship: physico-chemical or biochemical properties◦ Boiling point◦ Partition coefficients (LogP)◦ Receptor binding

Quantitative Structure-Activity Relationship: interaction with the biota◦ Toxicity◦ Metabolism

13

Activity dataActivity dataGarbage in, garbage out

Quality and quantity of data

◦ Suitable for purposes?

◦ Intrinsic variability of Y data (particularly for QSAR): examples later on

◦ Chemical domain covered with experimental data

◦ As much as you can expecially if using complex models

14

Data needData needData are one of the

pillars of the modelsThe goal is to

extract knowledge from these data

If they are too noisy it is not possible to extract this knowledge

Enough number of training data

Keep data variability low

Large number of compounds

Quality / Accuracy

Nr.

of

com

poun

ds

15

Modelling stepsModelling stepsData pre-processing

◦Scaling X block and transformation of Y block

Variable selection

Application of algorithms to search for the reationship

16

Scaling variables◦ making sure that each descriptor has an

equal chance of contributing to the overall analysis

◦ E.g.: autoscaling, range scaling

Y transformation

Data pre-processing (I)Data pre-processing (I)

17

No. TestSub Name_Jp CAS nr MolWeight ZM1V HNar MSD GMTIV SPI TI11 1-001 Trichlorfon 52-68-6 257,437 162,124 1,358 0,266 1336,074 9,072 -13,982

2 1-002 Dimethoate 60-51-5 229,257 148,198 1,485 0,324 1587,741 8,253 -16,476

3 1-003 Dichlorvos 62-73-7 220,976 172,519 1,451 0,326 1380,444 7,646 -13,908

4 1-004 Malathon 121-75-5 330,358 274,198 1,551 0,26 6524,185 13,757 -34,658

51-005 Methoprene

40596-69-8 310,471

224 1,562 0,334 9922 17,17 -57,135

6 1-006 Propylthiourea 927-67-3 118,201 50,444 1,448 0,431 235,667 4,166 -6,962

7 1-007 2-Butanone oxime 96-29-7 87,1204 72 1,385 0,414 224 3,484 -4,909

8 1-008 Dibromoacetic acid 631-64-1 217,844 86,134 1,286 0,38 204,949 3,786 -4,593

9 1-011 Bis(2-ethylhexyl)adipate 103-23-1 370,566 254 1,696 0,334 14655 17,037 -80,091

10 1-013 Thiram 137-26-8 240,433 87,778 1,44 0,338 829,445 9,053 -17,08

11 1-015 Stannane, tributylfluoro- 1983-10-4 309,051 88,008 1,6 0,314 1243,889 8,561 -22,079

121-016 Methomyl

16752-77-5 162,21

148,444 1,5 0,375 1128,333 6,542 -12,904

13 1-017 Aldicarb 116-06-3 190,263 158,444 1,485 0,351 1697 8,357 -17,609

14 1-018 Demeton-s-methyl 919-86-8 230,285 124,198 1,548 0,353 1163,185 7,554 -17,685

15 1-019 Citral 5392-40-5 152,233 106 1,535 0,387 1399 7,197 -16,06

16 1-020 Disulfiram 97-77-8 296,539 103,778 1,548 0,307 1801,444 11,43 -28,141

17 1-021 2-Ethyl-1,3-hexanediol 94-96-2 146,227 86 1,5 0,338 805 6,412 -11,897

18 1-022 Tributyl phosphate 126-73-8 266,314 183,309 1,659 0,311 3441,667 10,069 -32,437

20 1-024 Tris(2-chloroethyl)phosphate 115-96-8 285,49 170,124 1,6 0,314 1985,037 8,561 -22,079

22 1-026 Ethylene glycol 107-21-1 62,0678 58 1,333 0,527 139 1,893 -2,499

Data pre-processing (II)Data pre-processing (II)Variable pruning

◦ Detecting constant variables

◦ Detecting quasi-constant variables

It can distinguish between informative and non informative variables

◦ Detecting correlated variables

Variables can be grouped into correlation groups and the most correlated variable with the response is retained

◦ Variables with missing values

18

Variable selectionVariable selectionReducing dimensions, facilitating

data visualization and interpretation

Likely improving prediction performance

Hypothesis driven or statistically driven

19

Wrappers: utilizes the choice of prediction method to score subsets of features according to their predictive power;

Filters: a preprocessing step, independent of the choice of the predictor.

Variable selection Variable selection techniquestechniquesPrincipal component analysis (PCA)ClusteringSelf organizing maps (SOM)Stepwise procedures

• Forward selection: features are progressively incorporated into larger and larger subsets;

• Backward elimination: starting with the set of all features and progressively eliminates the least promising ones.

Genetic algorithmsVariable importance/sensitivity

20

Principal component Principal component analysisanalysisKeep only those components that

possess largest variationPC are orthogonal to each other

Loadings plot21

ClusterCluster analysis analysisProcess of putting objects into classes,

based on similarityDescriptors in the same cluster are

assume similar values for the molecules of the dataset

Many different methods and algorithms◦ different clustering methods will result in

different clusters, with different relationships between them

◦ different algorithms can be used to implement the same method (some may be more efficient than others)

22

Hierarchical and non-Hierarchical and non-hierarchicalhierarchicalA basic distinction is between

clustering methods that organise clusters hierarchically, and those that do not

3 42 5 6 7 81 3 42 5 6 7 81

23

Hierarchical Hierarchical agglomerativeagglomerativeThe hierarchy is built from the bottom

upwardsSeveral different methods and algorithmsBasic Lance-Williams algorithm (common

to all methods) starts with table of similarities between all pairs of items◦ at each step the most similar pair of

molecules (or previously-formed clusters) are merged together

◦ until everything is in one big cluster◦ methods differ in how they determine the

similarity between clusters

24

Hierarchical divisiveHierarchical divisiveThe hierarchy is built from the

top downwardsAt each step a cluster is chosen

to divide, until each cluster has only one member

Various ways of choosing next cluster to divide◦ one with most members◦ one with least similar pair of members◦ etc.

Various ways of dividing it25

Non-hierarchical methodsNon-hierarchical methodsUsually faster than hierarchical e.g.: Nearest neighbour methods

◦best known is example is Jarvis-Patrick method identify top k (e.g. 20) nearest

neighbours for each object

two objects join same cluster if they have at least kmin of their top k nearest neighbours in common

◦tends to produce a few large heterogeneous clusters and a lot of singletons (single-member clusters)

26

Self organizing mapsSelf organizing mapsA SOM is an unsupervised NN

condensing the input space into a low-dimensional representation

27

Genetic algorithmsGenetic algorithmsBased on the Darwinian evolutionary

theory◦ individuals in a population of models are crossed

over, mutated, then iteratively evaluated against a fitness function which gives a statistical evaluation of the model’s performances

28

Initialpopulation

Evaluation of individuals

Cross-over

MutationsIndividual selection

Fitness?

End

Y

N

10010111011011010101010101011110011100111010001010010010 10001010110010

11010111011111

Modelling approachesModelling approaches

SAR

Quantitative SAR

29

Categorical YClassification

Continuous YRegression

Modelling techniquesModelling techniquesMultiple Linear RegressionPLS…

Neural Networks

Classification treesDiscriminant analysisFuzzy classification…

30

Multiple RegressionMultiple Regression

Linear relationship between Y

and several Xi descriptors

Y = aX1 + bX2 + cXn + … + const.

Minimize error by least squares

May include polynomial terms

31

; (1)

)(

)ˆ(1

1

12

yy

yyR

i

n

i

ii

n

i

32

Partial Least SquarePartial Least Square

PLS similarly to PCA uses orthogonal PC of linearly correlated variables more closely related to the Y response

Scores t1&t2 projection

O = f(I)

I O

Neural networksNeural networksInspired by biological NNs are a set of connected

nonlinear elements making transformation of input

33

The problem of The problem of overfittingoverfitting

y = 0.979x + 0.344R² = 0.956

y = -0.062x4 + 1.293x3 - 9.472x2 + 29.24x - 27.37R² = 0.999 34

Solution: validationSolution: validation

35

Training prediction

Validation prediction

Complexity

Perf

orm

an

ces

Validation criteriaValidation criteriaInterna validation - robustness

◦Cross-validation (LOO, LSO)◦Bootstrap◦Y scrambling

External validation - prediction ability◦Test set representative of training

set◦Tropsha criteria

Applicability domain

36

Cross validationCross validationLeave One OutAll the data are used for fitting but one compoundPredict the excluded sampleRepeat it for all samplesCalculate Q2 or R2cv similarly to R2 on the basis of

these predictions

Problem: to optimistic if there are many samples

Leave Many OutUse larger groups to obtain a more realistic

outcome

37

BootstrappingBootstrappingBootstrapping simulates what happen by

randomly resampling the data set with n objects

K n-dimensional groups are generated by a randomly repeated some objects

The model obtained on the different sets is used to predict the values for the excluded sample

From each bootstrap sample the statistical parameter of interest is calculated

The estimation of accuracy is obtained by the average of all calculated statistics

38

Y-scramblingY-scramblingRandomply permutate Y responses while X

variables are kept in the same order for several times

39

Tropsha criteria*Tropsha criteria*

40

* A. Golbraikh, M. Shen, Z. Xiao, Y.D. Xiao, K.-H. Lee, A. Tropsha, Rational selection of training and test sets for the development of validated QSAR models, JCAMD, 17 (2003) 241-253.

a) Q2 > 0.5; b) R2 > 0.6;

c) (R2 - R20)/ R2 < 0.1 and 0.85 < k < 1.15 or

(R2 – R’20)/ R2 < 0.1 and 0.85 < k’ < 1.15

(k=slope of the regression line)

(R20 = R2 related to y=kx)

d) if (c) is not fulfilled, then | R20 – R’2

0| < 0.3

Applicability domainApplicability domain

The applicability domain of a (Q)SAR model is the response and chemical structure space in which the model makes predictions with a given reliability.*

41

* Current status of methods for defining the applicability domain of (quantitative) structure-activity relationships. ATLA, 33:1-19, 2005.

Applicability domainApplicability domain

42

Training data

Applicability domainApplicability domain

43

Training data

New compounds

AD assessmentAD assessment

Similarity measures:

Response range (span of activity data)

Chemometric treatment of the descriptor space

Fragment-based approaches

44

Chemometric MethodsChemometric Methods

Descriptor range-based

45

0

2

4

6

8

10

12

0 5 10 15 20

Descr. 1

De

sc

r. 2

Chemometric MethodsChemometric Methods

Descriptor range-based

Geometric methods

46

0

2

4

6

8

10

12

0 5 10 15 20

Descr. 1

De

sc

r. 2

Chemometric MethodsChemometric Methods

Descriptor range-based

Geometric methods

Distance-based

47

Chemometric MethodsChemometric Methods

Descriptor range-based

Geometric methods

Distance-based

Probability density

distribution

48

AMBIT softwareAMBIT software

http://ambit.acad.bg/main.php 49

AD assessmentAD assessment

Similarity measures:

Response range (span of activity data)

Chemometric treatment of the descriptor space

Fragment-based approaches

50

Example of AD Example of AD assessmentassessment

0

10

20

30

40

50

60

70

80

90

100Within 1 log unit

Within 2 log unit

Test set 1

Test set 2

0

10

20

30

40

50

60

70

80

90

100

% o

f com

poun

ds

Within 1 log unit

Within 2 log unit

% of all compounds in the test set predicted within one or two log unit without assessing the AD

51

Further aspects in ADFurther aspects in AD

52

Including the model’s characteristicsTerminal

nodeNode

assignment

Misclassification ratio

Training setValidation

setTest set

4 0 0.04 0.02 0

6 1 0.13 0.16 0.14

8 0 0.1 0.26 0.17

11 1 0.31* 0 0.25

12 0 0.05 0.14 0.25

14 1 0.47* 0.67* 0.25

15 0 0.14 0.19 0.13

18 1 0.32* 0.33* 0.55*

19 0 0.2 0 0

20 0 0.06 0.2 0.17

21 1 0.44* 0.2 0.17

OutlineOutlineHistoryQSAR/QSPR steps

◦(Descriptors)

◦Activity data

◦Modelling approaches

◦Validation (OECD principles)

QSPR (Phys-chem properties)QSAR (Biological activities)Example (DEMETRA)

53

Why?Why?

54

Large number of existing and new chemicals without a complete (eco)toxicological characterization

http://www.ewg.org/reports/skindeep/

Ingredient Search Results: [tocopheryl

acetate]

Ingredient Sample Product Categories Number of Products Ingredient Score (5=highest concern)

1. TOCOPHERYL ACETATE

Moisturizer , Facial Moisturizer/Treatment , Facial Cleanser , Body Wash/Cleanser , Lip Gloss

3144 0.6

Ingredient Categories: Low concern Moderate concern Higher concern

Cancer hazard Reproductive/developmental toxicity

Unsafe for use in cosmetics Illegal ingredients (EU)

Illegal ingredients (US) Unsafe in infant products

Potential for harmful impurities Ingredient(s) not disclosed on label

Sunburn/skin cancer risk Estrogenic chemicals and other endocrine disruptors

Irritants - eye, skin, or lungs Fragrance

Persistent/bioaccumulative Immune system toxicants (allergies, sensitization)

Penetration enhancers Safety limits on use/purity/manufacturing

Classified as toxic Potential for infectious disease risk

Hazards for occupational exposures Industry safety warnings

Illegal for use in food Illegal for use in drugs

Insufficient safety data Wildlife/environmental toxicity

Ingredient(s) not assessed for safety No safety information in 37 regulatory/toxicity data sources

Summary - health information

Constrains: time consuming, expensive,

ethical issues

REACHREACH

Enterprises that manufacture or import more than one tonne of a chemical substance per year would be required to register it in a central database

It is estimated that the testing of the approximately 30’000 existing substances would result in total costs of about 2,1 billion €, over the next 11 years

Promotion of non-animal testing

55

Registration, Evaluation and Authorisation of CHemicals

REACHREACH

56

Registration, Evaluation and Authorisation of CHemicals

Additional cost Use of (Q)SARs,

read-across 2.3 billion Euro Minimal

use

1.5 billion Euro Average use (likely scenario)

1.1 billion Euro Maximal use

Cost-saving potential: € 800-1130 million

Pedersen et al. (2003). Assessment of additional testing needs under REACH.

REACHREACH

57

Registration, Evaluation and Authorisation of CHemicals

Additional animalsUse of (Q)SARs, read-

across 3.9 million Minimal

use

2.6 million Average use (likely scenario)

2.1 million Maximal use

Animal-saving potential: 1.3-1.9 million animals

Van der Jagt et al. (2004). Alternative approaches can reduce the use of test animals under REACH.

OECD principles for QSAR OECD principles for QSAR validationvalidationEfforts to improve transparency and acceptability of in silico methods:

A defined endpoint

An unambiguous algorithm

A defined domain of applicability

Appropriate measures of goodness-of-fit, robustness and predictivity

A mechanistic interpretation, if possible

58

QSPRQSPRPhysico-chemical properties

◦Boiling point◦Solubility◦Partition coefficients◦Viscosity◦Hydrophobicity

Biochemical assays

59

Specific aspects of QSPRSpecific aspects of QSPRIn general you can expect to

obtain more precise models, and experience reduced experimental variability

Many properties important for drug design◦Biochemical assay – target property◦Bioavailability – LogP◦Side effects

Many others important for REACH60

QSARQSARBiological activities

◦Ecotoxicity◦Mammalian toxicity (as surrogate of

human health)◦Carcinogenicity & Mutagenicity◦…◦& many more

61

Specific aspcts of QSARSpecific aspcts of QSAR

Biological variability

Moles vs. wheight data

Role of LogP

Mechanistic interpretation

62

Biological variabilityBiological variabilityIntrinsic variability of

toxicological data (LC50)

63

Mole vs. Mole vs. wheight wheight unitsunits

64

Role of LogPRole of LogPUsed to model the penetration

into the phospholipidic membrane

Extreamly common for its easyness of interpretation

65

Role of LogPRole of LogPWhich is the your favourite option?

66

Tox = 1.32 • LogP + 0.23

Tox = 0.55 • des1 + 0.36 • des2 + 0.29 • des3 + 0.64 • des4 - 0.47 • des5 - 1.56 • des6 -

0.53 • des7 + 0.27 • des8 + 0.55 • des9 + 0.50 • des10 + 0.23

Role of LogPRole of LogP

67

KowWin (LogKow) Log P Calculation: SMILES : CC(C)(C)C(=O)C(Oc1ccc(Cl)cc1)n2cncn2 CHEM : Triadimefon MOL FOR: C14 H16 CL1 N3 O2 MOL WT : 293.76 -------+-----+--------------------------------------------+----------+--------- TYPE | NUM | LOGKOW v1.66 FRAGMENT DESCRIPTION | COEFF | VALUE -------+-----+--------------------------------------------+----------+--------- Frag | 3 | -CH3 [aliphatic carbon] | 0.5473 | 1.6419 Frag | 1 | -CH [aliphatic carbon] | 0.3614 | 0.3614 Frag | 8 | Aromatic Carbon | 0.2940 | 2.3520 Frag | 1 | -CL [chlorine, aromatic attach] | 0.6445 | 0.6445 Frag | 1 | -O- [oxygen, one aromatic attach] |-0.4664 | -0.4664 Frag | 1 | -C(=O)- [carbonyl, aliphatic attach] |-1.5586 | -1.5586 Frag | 3 | Aromatic Nitrogen [5-member ring] |-0.5262 | -1.5786 Frag | 1 | -tert Carbon [3 or more carbon attach] | 0.2676 | 0.2676 Factor| 1 | -N-C-O- structure correction | 0.5494 | 0.5494 Factor| 1 | -C-CO-C-O- structure correction | 0.5000 | 0.5000 Const | | Equation Constant | | 0.2290 -------+-----+--------------------------------------------+----------+--------- Log Kow = 2.9422

LogKow Estimated Log P: 2.94

From LogPFrom LogP

68

descriptorsdescriptors

structure

logP

activity

descriptors

structure activity

To direct descriptors

MechanisticMechanistic interpretationinterpretationA priori (experimentally

determined – even more complex that the studied endpoint itself) or postulated or a posteriori

Different classification schemes for MOA exist (narcosis, specific reactive modes)

69

Global models …Global models …

70

Training setTraining set n = 422 d = 5 R2 % = 69,9 %Rcv2 % = 68,0 % RMS = 0,77

Test setTest set n = 141 R2 = 71,7 RMS = 0,70

- Log (LC50) for training and test set

Observed

-4 -2 0 2 4 6

Pred

icte

d

-4

-2

0

2

4

6 Training setTest set

N-Vinylcarbazole

Acrolein

2-propyn-1-ol

2-propen-1-olMalononitrile

Global models …Global models …

71

nn 563 563

dd

Log P, ELog P, ELUMOLUMO, MW,

Kier&Hall (order 0), Molecular surface area

Log P, Log P,

EELUMOLUMO

RR22 71.1 69.5

QQ22 70.7 69.3

RMSRMS 0.74 0.76

vs mechanistic models …vs mechanistic models …

72

-Log (LC50) for reactive compounds

Observed

0 2 4 6

Pre

dict

ed

0

2

4

6

n d R2% Rcv2% RMS

141 2 58.6 58.2 0.83

MOA Class n d R2% Rcv2%

Narcosis I

Narcosis II

Narcosis III

238

38

26

2

3

4

90.1

82.9

91.7

89.9

81.1

90.6

OutlineOutlineHistoryQSAR/QSPR steps

◦(Descriptors)

◦Activity data

◦Modelling approaches

◦Validation (OECD principles)

QSPR (Phys-chem properties)QSAR (Biological activities)Example (DEMETRA)

73

Activity data collectionActivity data collection Identification of the endpoints that can

mostly benefit for QSAR Costs, test severity,

feasibility, etc…

Identification of data sources Quality, guidelines, protocols

Refinement of the data Multiple sources comparison,

precautionary selection

74

TROUT(282)

WATER FLEA

(263) ORAL QUAIL (116)

DIETARY QUAIL (123) BEE

(105)

WHOLE

DATASET

(398)

Trout

Daphnia

Oral quail

Dietary quail

Bee

• Individual models

Linear models, ANN models

• Hybrid system

Combining

model results

Modelling processModelling process

75

ValidationValidationValidation of the hybrid model for

Daphnia with new data subsequenly identified in literature◦Real “blind” test set

Comparison with Expert systems◦ECOSAR◦Topkat

76

DEMETRA Hs - training DEMETRA Hs - training setset

77

-4

-2

0

2

4

6

-4 -2 0 2 4 6

experimental value [-log(mg/l)]

pre

dic

ted

val

ue

[-lo

g(m

g/l)

]Daphnia Magna

TRAINING SETNC = 193

R2 = 0.80

DEMETRA Hs - DEMETRA Hs - results on results on teststests

78

Daphnia Magna

TEST SETSEPA test set

NC = 36

R2 = 0.80

-4

-2

0

2

4

6

-4 -2 0 2 4 6

experimental value [-log(mg/l)]

pre

dic

ted

val

ue

[-lo

g(m

g/l)

]

D-BBA test set

NC = 101

R2 = 0.70

US EPA ECOSAR US EPA ECOSAR predictionspredictions

79

Daphnia Magna

ECOSAR – tutti i datiNC = 432

R2 = 0.20

-4

-2

0

2

4

6

-4 -2 0 2 4 6

experimental value [-log(mg/l)]

pre

dic

ted

val

ue

[-lo

g(m

g/l)

]

Topkat predictionsTopkat predictions

80

NC = 176

NC (training test) = 31

R2 = 0.20

-4

-2

0

2

4

6

-4 -2 0 2 4 6

experimental value [-log(mg/l)]

pre

dic

ted

val

ue

[-lo

g(m

g/l)

]