Feature Extraction Methods in Quantitative Structure–Activity Relationship Modeling: A Comparative Study

Computational approaches for synthesizing new chemical compounds have resulted in a major explosion of chemical data in the field of drug discovery. The quantitative structure-activity relationship (QSAR) is a widely used classification and regression method used to represent the relationship between a chemical structure and its activities. This research focuses on the effect of dimensionality-reduction techniques on a high-dimensional QSAR dataset. Because of the multi-dimensional nature of QSAR, dimensionality-reduction techniques have become an integral part of its modeling process. Principal component analysis (PCA) is a feature extraction technique with several applications in exploratory data analysis, visualization and dimensionality reduction. However, linear PCA is inadequate to handle the complex structure of QSAR data. In light of the wide array of current feature-extraction techniques, we perform a comparative empirical study to investigate five feature-extraction techniques: PCA, kernel PCA, deep generalized autoencoder (dGAE), Gaussian random projection (GRP), and sparse random projection (SRP). The experiments are performed on a high-dimensional QSAR dataset, which comprises 6394 features. The transformed low-dimensional dataset is inputted into a deep learning classification model to predict a QSAR biological activity. Three approaches are adopted to validate and measure the proposed techniques: (i) comparing the performance of the classification models, (ii) visualizing the relationship (correlation) between features in the low-dimension Euclidean space, and (iii) validating the proposed techniques using an external dataset. To the best of our knowledge, this study is the first to investigate and compare the aforementioned feature-extraction techniques in QSAR modeling context. The results obtained provide invaluable insights regarding the behavior of different techniques with both negative and positive classes. With linear PCA as a baseline, we prove that the investigated techniques substantially outperform the baseline in multiple accuracy measures and demonstrate useful ways of extracting significant features.