1. QSAR MODELLING OF THE AROMATIC AMINES MUTAGENICITY BY GENETIC ALGORITHM - VARIABLE SUBSET SELECTION
M.Pavan, P.Gramatica, F.Consolaro, V.Consonni, R.Todeschini
QSAR Research Unit, Dept. of Structural and Functional Biology, University of Insubria, Varese, Italy
Web:
INTRODUCTION
Aromatic and heteroaromatic amines are widespread chemicals of considerable industrial and environmental relevance as they are carcinogenic for human beings. QSAR studies have been used to develop models to estimate and to predict mutagenicity by relating it to chemical structure. In mutagenicity QSAR applications, the investigators focus on either the molecular determinants that discriminate between active and inactive chemicals, or the modulators of the relative potency of the active chemicals. The development of a model to predict mutagenicity necessitates a test system capable of providing reproducible and quantitative estimates of toxic activity; the most widely used is a bacterial test, based on the Salmonella typhimurium strains (TA98  frameshift mutation; TA100  base-substitution mutation), introduced by Ames.
The data set is constituted by 146 aromatic and heteroaromatic amines collected by Debnath1; mutagenicity data are expressed as the mutation rate in log (revertants/nmol).
[1]A.K. Debnath et all. A QSAR investigation of the Role of Hydrophobicity in Regulating Mutagenicity in the Ames Test: 1. Mutagenicity of Aromatic and Heteroaromatic Amines in Salmonella typhimurium TA98 and TA100. Environmental and Molecular Mutagenesis 19, (1992).
MOLECULAR DESCRIPTORS
The molecular structure has been represented by a wide set of 657 molecular descriptors calculated by the software DRAGON2:
constitutional descriptors (56) BCUT descriptors (7)
walk counts (20) 2D autocorrelation descriptors (96)
Galvez index (21) aromaticity descriptors (4)
charge descriptors (7) geometrical descriptors (18)
molecular profiles (40) WHIM descriptors (99) 3
3D-MoRSE descriptors (160) empirical descriptors (3)
topological descriptors (69)
[2]R.Todeschini and V.Consonni - DRAGON - Software for the calculation
of molecular descriptors, version 1.0 for Windows,(2000), Milano Chemometric and QSAR Research Group.Free download available at:
[3]R.Todeschini and P.Gramatica, 3D-modelling and prediction by WHIM
descriptors. Part 5. Theory development and chemical meaning of the
WHIM descriptors, Quant.Struct.-Act.Relat., 16 (1997)
Training set
Internal
validation
Test set
External
Models
Predictions
Q2LOO
Q2LMO Zr
ERcv
Q2ext
ERext
TRAINING SET SELECTION
In order to have knowledge of the predictive capability of the models both internal and external validations were performed. An experimental design based on the Todeschini-Marengo algorithm was used to select the most representative training set of amines: models were developed on the selected training set and predictions were made for the molecules excluded from the model generation step (test set).
Molecular descriptors
Experimental responses
DATASET: amines
657 molecular descriptors
2 responses - TA98
- TA100
Training set
Test set
Variable subset
selection
Regression
models
Classification
CART
K-NN
RDA
CP-ANN
OLS
REGRESSION MODELS
The mutagenicity potency has been modelled by Ordinary Least Squares (OLS) method using a selected subset starting from 657 different molecular descriptors; the selection of the best subset of variables has been realised by Genetic Algorithm (GA-VSS). The obtained models have been validated by leave-one-out (Q2LOO), leave-more-out (Q2LMO), y-scrambling (Zr) and an external test set (Q2ext) and show satisfactory predictive performances, considering the uncertainty of the biological end-points.
TA98
TA100
LogTA98= MWC MATS7m+2.44Mor27u+1.12Mor15m
LogTA100= nN-16.34ATS2e+12.43ATS4p-1.66GATS4p
Training set = 60 compounds
Test set = 39 comp.
Training set = 46 compounds
Test set = 30 comp.
n.variables Q2LOO Q2LMO Q2ext R2
n.variables Q2LOO Q2LMO Q2ext R2
nN = number of Nitrogen atoms
ATS2e = Broto-Moreau autocorrelation of a topological structure - lag 2 / weighted by atomic Sandreson electronegativities
ATS4p = Broto-Moreau autocorrelation of a topological structure - lag 4 / weighted by atomic polarizabilities
GATS4p = Geary autocorrelation - lag 4 / weighted by atomic polarizabilities
MWC07 = number walk count of order 07
MATS7m = Moran autocorrelation - lag 7 / weighted by atomic masses
Mor27u = 3D-MoRSE - signal 27 / unweighted
Mor15m = 3D-MoRSE - signal 15 / weighted by atomic masses
CLASSIFICATION MODELS
Some classification methods (CART, K-NN, RDA and CP-ANN) have been applied to this data set in order to distinguish between activity classes. The selection of the best subset of variables has been realised by the experimental design based on the Todeschini - Marengo algorithm. The models have been validated internally (ER) and externally (ERext). The classification models for TA100 have showed a predictive power worse than the TA98 ones and thus they are not presented here.
CONCLUSIONS
frameshift mutation
STRAIN TA98
molecular dimension
molecular branching
TA98
CART (classification and regression tree)
aromatic amines
intercalary agents
BEPp1 < 3.77 2 1
NOMER% ER% ERext%
Training set = 55 compounds
Test set = 60 comp.
STRAIN TA100
base-substitution mutation
molecular dimension
electronic properties
aromatic amines
complex base-pair substitution mutation
1 = mutagenic compounds
2 = non mutagenic compounds
BEPp1 = positive eigenvalue n.1 / weighted by atomic polarizabilities