Structural and functional interpretation of QSAR models

P.G. Polishchuk1, T.M. Khristova1,2, L.N. Ognichenko1,

A.P. Kosinskaya1, A.A. Varnek2, V.E. Kuz’min1

1 A.V. Bogatsky Physico-Chemical Institute of NAS of Ukraine, Odessa, Ukraine2 Strasbourg University, Strasbourg, France

31 August – 4 September, 2014St.Petersburg, Russia

QSAR interpretation: interpretability vs. complexity

2

Mo

del

inte

rpre

tab

ility

Model complexity

Popular misbelief

DTMLR

PLS

ANN

kNNRF

SVM

models ensembles

QSAR interpretation approaches

Model-specific approaches:

Rule-based (Decision tree)

Regression coefficients (MLR, PLS)

Latent variables (PLS)

Weights and biases (ANN)

…

Model-independent approaches:

Local gradients or partial derivatives

3i

iiii

x

)xf(x)f(x

ΔC

Model

QSAR interpretation: common workflow

Variables contributions

Structure-property

relationship

f(x)Var_1 Var_2

Mol_1 -0.23 1.82

Mol_2 2.36 1.27

Mol_3 5.01 2.30

Mol_4 0.69 -0.58

4

Matched molecular pairs & molecular transformations

logS = -3.18 logS = -0.60

H → OH

ΔlogS = 2.58

logS = -2.21 logS = -0.62

ΔlogS = 1.59

5Leech A.G. et al, J. Med. Chem. 2006, 49, 6672-6682Sheridan R.P. et al., J. Chem. Inf. Model. 2006, 46, 180-192

Exemplified dataset

6

Structural QSAR interpretation

- =

logSpred = -1.55 logSpred = -1.61 ΔlogSpred = 0.06

- =

logSpred = -1.55 logSpred = -1.35 ΔlogSpred = -0.207

Polishchuk P.G. et al. Molecular Informatics 2013,32, 843-853

Structural QSAR interpretation

logSpred = -1.93 logSpred = -4.32 ΔlogSpred = 2.39

- =

8Polishchuk P.G. et al. Molecular Informatics 2013,32, 843-853

Atom-property labeling

Simplex generation example

Kuz’min, V. E. et al, Journal of Molecular Modelling 2005, 11, 457-467.Kuz’min, V. E. et al, Journal of Computer-Aided Molecular Design 2008, 22, 403-421.

Simplex representation of molecular structure (SiRMS)

9

Examples of structural interpretation

10

Solubility (1033 compounds)

11

Endpoint ModelSiRMS Dragon

R2CV RMSE R2

CV RMSE

Solubility,logS

PLS 0.84 0.82 0.91 0.60RF 0.88 0.71 0.91 0.62

SVM 0.87 0.72 0.92 0.59

5-fold external cross validation results

Mutagenicity (Ames, 4361 compounds)

12

Descriptors Algorithm Balanced Accuracy

SiRMSRF 0.817

SVM 0.800

DragonRF 0.816

SVM 0.793

5-fold external cross validation results

Combined contribution (effect) of fragments

13

RF+SiRMS

Contribution = 0(non-mutagen)

Contribution = 0(non-mutagen)

Contribution = 1(mutagen)

Questions

14

How does the fragmentinfluence the property?

WHY?!

pIC50 = f(A1, A2, A3) = x pIC50 = f(B1, B2, B3) = y

Functional interpretation of QSAR models

15

1, 2, 3 – groups of descriptors represented different physico-chemical factors(charge, H-bonding, etc) of compound A and B.

A B

- =

pIC50 = f(A) = x pIC50 = f(B) = y Contribution(C) = x - y

Structural interpretation

C

pIC50 = f(A1, A2, A3) = x pIC50 = f(A1, A2, B3) = y Contribution3(C) = x - y

Functional interpretation

A B

- =C

Antagonists of fibrinogen receptor

16

Arg-mimetic Linker Asp-mimetic

Frag

men

t ex

amp

les

Antagonists of fibrinogen receptor: dataset

17

Arg-Gly-Asp

338 compounds

Antagonists of fibrinogen receptor: models

18

Algorithm R2 RMSE

RF 0.72 0.80

SVM (RBF kernel) 0.69 0.84

SVM (linear) 0.67 0.88

PLS 0.67 0.87

5-fold external cross validation results

Structural interpretation

19

Arg

-mim

eti

cLi

nke

rA

sp-m

ime

tic

Functional interpretation

20

Arg

-mim

eti

cLi

nke

rA

sp-m

ime

tic

Functional interpretation of RF model

21

Arg214

Mg2+

Asp224

Ser225

0.070.74

0.341.77

0.051.03

0.420.6

0.051.09

0.461.02

0.040.29

0.140.81

electrostaticH-bondinghydrophobicitypolarizability

Phe160Tyr190

RF model

Functional interpretation of SVM model

22

Mg2+

0.210.71

0.220.72

0.080.51

0.160.56

0.110.46

0.220.81

0.10.130.14

0.39

electrostaticH-bondinghydrophobicitypolarizability

Arg214

Asp224

Ser225

Phe160Tyr190

SVM-RBF model

Conclusions

23

Almost any QSAR model can be interpreted using the proposed schemes.

Results of structural and functional interpretation obtained from different models are well correlated.

Structural interpretation allows fragments ranking and finding potential alerts.

Functional interpretation may provide a guess which factors are dominated and influence on the investigated property.

A.V. Bogatsky Physico-Chemical Institute, Chemoinformatic group:http://qsar4u.com

SiRMS project on GitHub:https://github.com/DrrDom/sirms

Online structural interpretation tool:http://www.physchem.od.ua/compute

Related projects

24

http://qsar4u.com/https://github.com/DrrDom/sirmshttp://www.physchem.od.ua/compute