Computational approaches can be classified into structure-based drug design (SBDD), ligand-based drug design (LBDD) and sequence-based approaches (Ou-Yang et al., 2012). In the presence of experi­mentally determined structure of target molecule, molecular docking and de novo drug design can be used in the SBDD procedures. In situations in which three dimensional (3D) structure of target mol­ecule is not available, LBDD methods are applied. Quantitative structure-activity relationship (QSAR), molecular field analysis, pharmacophore modeling, and 2D/3D similarity assessment are examples of these methods. In the absence of information regarding target and ligand molecules, sequence-based approaches based on bioinformatics tools are utilized to ultimately find a solution to the drug develop­ment problem which may fit in one of the two previously mentioned categories.

QSAR methodology covers data analysis and statistical methods for developing models with capability of predicting the biological and toxicological activities of compounds based on their structures. It was first introduced by Corwin H. Hansch (Hansch, Maloney, Fujita, & Muir, 1962). The basic principle of the technique relies on building mathematical relationship between biologi­cal/toxicological activity data of a series of compounds and their structures defined by molecular descriptors. Molecular descriptors are numerical representations of molecular structures derived from steric, electronic, constitutional, topological, and geometrical properties of the molecules. Descriptors used in QSAR hold information about a molecule either as a whole or distinct parts of it. These descriptors can be experimentally measured or calculated using theoretical methods. Having descriptors in hand, those with biological relevance should be selected. Different descriptor selection methods, aimed to determine the best descriptor subset from a large pool of descriptors, have been proposed to reduce the number of descriptors present in the final mathematical QSAR model for explaining the biological activities of the studied molecules. These methods can be classified in three common categories including classical, artificial intelligence based, and miscel­laneous methods that are necessary for generalization of predicted QSAR model through training steps. Multiple linear regression (MLR) analysis and partial least squares (PLS) are the examples of linear methods whereas artificial neural networks (ANNs) and the k-nearest neighbours (k-NNs) are few examples for non-linear methods. Generally, linear models predict the biological activity in a linear function and these models are also easy to interpret compared with non-linear models. Although, the classical methods can model the linear relationships between structural properties and biological activities, however, they are not effective in developing models where the relationships are non-linear. Therefore, in these cases, the descriptor selection methods based on artificial intel­ligence can easily map existing non-linear relations between structural properties represented by descriptors and their corresponding biological activity faster and with higher reliability than classical methods because of their non-linear architecture. As QSAR methods are predictive modeling tools, the important criterion for evaluating their reliability and validity is to assess the capability of the models to predict not only the biological activity of the molecules included in training set (internal validation) but also for the molecules not involved in generating the model (external validation). Leave-one-out (LOO) and leave-group-out (LGO) are commonly utilized methods for internal cross validation whereas for external validation method, the predictivity of the QSAR model is judged based on a separate set of molecules called test set which are not included in model development in any stage. QSAR has served as a valuable predictive tool for identifying hit compounds, predict­ing the biological/toxicological activity of untested pharmaceuticals as well as optimizing the lead compounds for improving the biological activity (shown in Figure 1).

logical activity is important while developing the QSAR models. Full spectrum of activity is required for a data set and should range at least 3 to 4 orders of magnitude difference between the most and the least potent compounds. The larger the range, the more reliable QSAR model can be achieved (Scior et al., 2009). The compounds with poor activity should be incorporated in developing a predictive QSAR model. However, remotely isolated rare activities may be problematic by causing biases in the developed model. Since, the data for biological activity are crucial in generating accurate and valid QSAR model, therefore too much attention should be paid as they are not always as good as they seem at first sight.

There are several different ways of representing chemical structures including 2D and 3D representa­tions. 2D structures specify elements, connectivity, and bond order whereas Cartesian coordinates for each atom is determined in 3D formats. Encoding approaches for 2D representations are line notations, connection tables and molecular graphs. Line notations are appeared as a sequence of atoms connected by compact linear string of alphanumeric symbols. The most commonly used line notations are described as follows. Simplified Molecular Input Line Entry System (SMILES; see Figure 2) is widely used for­mat of notation implemented by Daylight Chemical Information Systems (Santa Fe, NM) (Weininger, 1988; Weininger, Weininger, & Weininger, 1989). SYBYL Line Notation (SLN) developed by Tripos Inc. provides richer set of information for molecular queries, stereochemistry, and reactions (Figure 2). The other line notation introduced recently is InChI stands for IUPAC international chemical identifier. This format of representation provides a unique line notation consisting of six layers, each identifying a distinctive structural property about the molecule. The layers are main layer, a charge layer, stereochemi­cal layer, an isotopic layer, a fixed-H, and a reconnected layer. The presence of main layer, which speci­fies atom and bond types, is required but the others are optional (Warr, 2011; Young, 2009). The other molecular line notations are Wiswesser Line-Formula Notation (WLN) and representation of structure diagram arranged linearly (ROSDAL). Figure 2 shows different line notation formats for phenylalanine as an example (Warr, 2011).

The other method for representing the chemical structures in 2D format is the molecular graph. Molecular graphs are derived from the graph theory; a theory not only used in chemistry but also in computer and related sciences. This theory is a branch of mathematics that defines the sets of objects by nodes and the corresponding relationships between them called edges. Atoms and bonds are considered as nodes and edges, respectively, in a molecular graph. For a molecular graph some properties like atom type (for nodes) and bond order (for edges) can be represented. Generally, the molecular graph shows the topology of the molecule and the way the atoms are connected (Figure 4).