A Comparison of Dimensionality Reduction Techniques in Virtual Screening

Abstract

Most of the screening methods have always struggled to deal with the high dimensionality of data in virtual screening task. One of the most commonly used techniques to reduce the high dimensional data is principal component analysis (PCA). PCA and its variants have been introduced and re-introduced to solve the problems in particular tasks in real world applications. In this paper, PCA and four variants of it are compared and analyzed together in virtual screening task in particular using fingerprint representation. Fingerprint is one of the most regularly used descriptors in virtual screening task. None of these methods have never been compared and studied together with high dimensional and binary-valued data elsewhere. The results show superiority of the variants of PCA to PCA on the most heterogeneous classes, while the methods are competitive to PCA on the homogeneous classes. Supervised PCA is found to be the best technique and is competitive to Fisher discriminant analysis. It should be noted that Fisher discriminant analysis uses all the provided information while Supervised PCA uses only few components.