*2.2.2 CoMFA studies*

*2.1.2 Molecular descriptors generation*

*Sino-Nasal and Olfactory System Disorders*

*2.1.3 Statistical analysis*

level (*P*-value) [19].

**2.2 3D-QSRR study**

*2.2.1 Minimization and alignment*

performed with 20 cycles.

**Software Descriptors**

**Table 2.**

**158**

G (kJ mol<sup>1</sup>

Twenty-eight molecular descriptors were calculated using ACD/ChemSketch

To explain the structure-property relationship, 28 descriptors are calculated for the 29 molecules using the ChemOffice and ChemSketch software, and they were subjected to a stepwise multiple linear regression (MLR) available in the SPSS software [18]. The stepwise MLR was generated to predict retention property values Log(LRI). Equation was justified by the correlation coefficient (r), the root mean square of the errors (RMSE), the Fishers F-statistic (F), and the significance

The final stage of this 2D-QSRR analysis consists of statistical validation in order to assess the significance of the model and hence its ability to predict property of other compounds. In this chapter, the model was validated internally by the crossvalidation test. The cross validations are statistical techniques in which different proportions of chemicals are iteratively held out from the training set used for model development. In this chapter, the leave-one-out procedure is used; this process sequentially removes one compound from the training set containing 24 compounds. A 2D-QSRR model is created on a "23" set of molecules, and the molecule removed is predicted by the constructed model. This process is repeated "24" times in order to predict the retention property of all compounds [20].

Chemical structures of studied compounds were sketched with sketch module in SYBYL [21] and minimized using Tripos force field [22] with the Gasteiger-Hückel charges [23] and conjugated gradient method, and gradient convergence criteria of 0.01 kcal/mol. Simulated annealing on the energy minimized structures was

Molecular alignment is one of the most sensitive parameters in 3D-QSRR methods. In this work, all studied compounds were aligned on the common core (compound no. 1), using the simple alignment method in Sybyl [24]. Compound

ChemOffice Melting point T (Kelvin); molecular weight MW (g/mol); critical temperature CT

ChemSketch Percent ratios of nitrogen, hydrogen, oxygen, and carbon atoms (H%; O%; C%); surface

); critical pressure CP (Bar); Connolly solvent-excluded volume V (A°)<sup>3</sup>

shape coefficient I; total connectivity TC; Log P; number of rotatable bonds NRB; winner index (W); number of H-bond acceptors (NHA); molecular topological index MTI; number of H-bond donors (NHD); partition coefficient PC; Balaban index (J);

); boiling point TB (Kelvin); Gibbs free energy

; total valence connectivity TVC;

;

(Kelvin); heat of formation H° (kJ mol<sup>1</sup>

sum of valence degrees SVD

Henry's law constant KH; polar surface area PSA (A°)<sup>2</sup>

tension γ (dyne/cm); index of refraction (n); density (d)

*Descriptors selected and software packages used in the calculation of descriptors.*

and ChemOffice programs [15, 16] to predict the correlation between these descriptors and the retention property of studied compounds and to develop a linear model [17]. The descriptors used in this study are displayed in **Table 2**.

> Based on the molecular alignment, CoMFA studies were performed to analyze the specific contributions of steric and electrostatic effects. These interactions were calculated using the Tripos force field with a distance-dependent dielectric constant at all interactions in a regularly spaced (2 Å) grid taking a sp3 carbon atom as steric probe and a +1 charge as electrostatic probe. The cutoff was set to 30 kcal/mol [25]. With standard options for scaling of variables, the regression analysis was carried out using the fully cross-validated partial least squares (PLS) method (leave one out) [26]. The final model that is non–cross-validated conventional analysis was developed with the optimum number of components to yield a non–cross-validated r2 value.

### *2.2.3 Partial least squares analysis (PLS) and validation*

The 3D-QSRR models were generated using a training set of 24 molecules. Predictive power of the resulting models was evaluated using a test set of five molecules (**Table 1**). The test compounds have been selected randomly. PLS analysis used to construct the 3D-QSRR models is an extension of multiple regression analysis in which the initial variables are replaced by optimum number of components of their linear combinations. PLS statistical method with leave-one-out (LOO) cross-validation procedure was used in this work to determine the optimal numbers of components considering cross-validated coefficient rCV for the training set of 24 molecules. The external validation of created models was determined using five compounds (test set). The final analysis (non-cross-validated analysis) was carried out using the optimum number of components obtained from the cross-validation analysis to get correlation coefficient r2 [27, 28].
