Descriptors and their selection methods in QSAR analysis: paradigm for drug design

Highlights

• A few newly introduced molecular descriptors were discussed.

• Various computational approaches to calculate the descriptors are listed.

• We described several methods for descriptors selection for building high predictive QSAR models.

• Advantage and disadvantages of selection methods were also addressed.

• Studies successfully applied the descriptors and their selection methods were also addressed.

The screening of chemical libraries with traditional methods, such as high-throughput screening (HTS), is expensive and time consuming. Quantitative structure–activity relation (QSAR) modeling is an alternative method that can assist in the selection of lead molecules by using the information from reference active and inactive compounds. This approach requires good molecular descriptors that are representative of the molecular features responsible for the relevant molecular activity. The usefulness of these descriptors in QSAR studies has been extensively demonstrated, and they have also been used as a measure of structural similarity or diversity. In this review, we provide a brief explanation of descriptors and the selection approaches most commonly used in QSAR experiments. In addition, some studies have also demonstrated the positive influence of features selection for any drug development model.

Introduction

Differentiating between drug-like from non-drug-like molecules is essential to reduce the cost associated with failed drug development. Various in silico approaches have shown the potential for screening chemical databases against the desired biological targets for the development of new potential leads [1]. Among them, ligand-based virtual screening has become popular because of its ability to screen millions of molecules rapidly from available chemical databases [2]. QSAR modeling is an important approach in drug discovery that correlates molecular structure with biological and pharmaceutical activities [3]. Such 2D methods rely on the calculation and comparison of molecular properties with the aim of identifying molecules that are similar with respect to the query molecule. Compared with 3D (or structure-based) methods, 2D approaches require substantially lower calculation times and, therefore, are mostly used as preliminary filters to reduce the number of compounds that can be used for further screening in later stages of drug development [4]. These 2D approaches are widely used in academia, industry, and research institutions worldwide. For the development of a QSAR model, one should consider it in terms of (i) the fundamental chemistry of the set of analogs, including any outliers; (ii) quantitatively correlating and summarizing the relations between chemical structure alterations and relevant changes in biological endpoint to determine the chemical properties that are the most likely determinants of the biological activities of the drug candidate; (iii) optimizing the existing leads to improve their biological activities; and (iv) predicting the biological activities of untried compounds.

Different QSAR approaches have been developed over the past few decades 5, 6, 7, 8. These approaches can determine the reliable relations between variations in the values of calculated descriptors and the biological activity for a series of chemical molecules, so that they can be used for predicting the activity of untried or newly synthesized compound(s). The chemical structures used in QSAR model building are encoded by a substantial number of molecular descriptors. The model is built by using only a few descriptors that are valid for closely related compounds. Most of the learning algorithms become computationally intractable when numbers of features are large, such as in the training algorithm and production steps. High-throughput data used in statistical modeling pose a challenge to accurate prediction. Given the large amount of inherent noise and variation in samples and their high dimensionality, there is the risk of overfitting [9]. Thus, there is a need for descriptor selection to improve model performance and avoid overfitting.

Descriptor selection methods provide a way of reducing computation time, improving prediction performance, and providing a better understanding of the data in machine learning. Descriptor selection is an important step for several reasons [10], including: (i) using only a few descriptors increases the interpretability and understanding of resulting models; (ii) It can reduce the risk of overfitting from noisy redundant molecular descriptors; (iii) it can provide faster and cost-effective models; and (4) it removes the activity cliff. However, noisy, redundant, or irrelevant descriptors should be removed in a way that the dimension of the input space is reduced without any loss of significant information [3]. In this review, we provide an update on, and a brief explanation of, commonly used descriptors, with a particular emphasis on their selection approaches for the development of more reliable, predictable, and generalized QSAR models.