Background: Microtubules are dynamic filamentous cytoskeletal proteins which have used widely in
cancer chemotherapy. Generally, the action of these compounds depends on binding to the tubulin proteins which are
α, β -heterodimers that form the core of the microtubules. Arylthioindoles (ATIs) inhibit tubulin polymerization by
binding to the colchicine site, inhibiting the binding of colchicine to tubulin. QSAR modeling play an important role
in modern medicinal chemistry, presenting a unique potential for transforming the early phases of drug research, in
terms of time and money savings. In this study we are intending to achieve some useful information about the
structural properties of mentioned ATIs using the QSAR modeling method. In this research, the QSAR modeling has
been employed to evaluate the efficacy of ATIs derivatives in the inhibition of tubulin polymerization. The best
multivariate linear model for each training and test data subset calculated using both genetic algorithm-multiple
linear regressions (GA-MLR) and stepwise-multiple linear regressions (S-MLR) methods. The best RS trial set has
been selected by comparing the two different QSAR modeling results in all train and test subsets. Finally, the S-MLR
and GA-MLR results were compared for finding the best random selection (RS) subset.
Methods: The molecular structures and experimental activity values for inhibition of tubulin polymerization obtained
from the literature. Here, using Dragon package over 1185 molecular descriptors such as RDF, GETAWAY, WHIM,
3D-MoRSE descriptors, and functional group descriptors derived for proper characterizing the arylthioindole
derivatives structures. These molecular descriptors reduced to 437 one by eliminating the constant variables,
collinear descriptors and the variables having low correlation with response. In fact in order to obtain the best RS
subsets in the mentioned method, the random samplings of the training subset (80% of data) were carried out 20
times and the remaining molecules (20% of data) considered as external validation set. The QSAR models were
constructed using stepwise-MLR and GA-MLR.
Results: In order to select important variables the GA-MLR and S-MLR methods applied on the training subsets
after data splitting. The best MLR models (models with high fitness function values) with four, five and six variables
built to obtain the best QSAR model. The best MLR model had four parameters in both GA-MLR and S-MLR
methods. The best significant relationships, using comparison of Q2 of models, for the IC50 values of tubulin assembly in
the models obtained in S-MLR and GA-MLR methods are presented for all of the random sets and SOM set.
Conclusion: The result of external validation indicates that in the study of inhibition of tubulin polymerization, the Q2test values are 0.7242 and 0.3215 for GA-MLR and S-MLR methods respectively. Considering the results which
obtained in both methods it can be concluded that in variable selecting the GA-MLR method is more powerful than
the S-MLR method.