Virtual Screening Using Binary Kernel Discrimination:  Analysis of Pesticide Data

Abstract

This paper discusses the use of binary kernel discrimination (BKD) for identifying potential active compounds in lead-discovery programs. BKD was compared with established virtual screening methods in a series of experiments using pesticide data from the Syngenta corporate database. It was found to be superior to methods based on similarity searching and substructural analysis but inferior to a support vector machine. Similar conclusions resulted from application of the methods to a pesticide data set for which categorical activity data were available.

Evaluation of machine-learning methods for ligand-based virtual screening

“…This fingerprint type encodes circular substructures centred on each non-hydrogen atom in a molecule by a string of extended connectivity values that are calculated using a modification of the Morgan Algorithm. The results obtained using this representation are shown in the right-hand column of Table 1, where it will be seen that a consistently high level of performance is obtained, as has been noted in our previous experiments using BKD [38,40,41]. In fact, there is very little difference between the results from BKD and CKD: for example, the average BKD recall for the datasets in Table 1(a) is 79.7%, just slightly more than the 78.1% for CKD.…”

“…Harper et al demonstrated that there was one kernel method, appropriately named binary kernel discrimination (BKD), that could be used with such data, and described the successful application of BKD to several pharmaceutical datasets [25]. Spurred by this study, we have reported the application of BKD to both pharmaceutical and agrochemical data and investigated some of the inherent characteristics of the approach [38][39][40][41]; here, we describe a kernel discrimination method for the analysis of non-binary molecular representations.…”

Support-Vector-Machine-Based Ranking Significantly Improves the Effectiveness of Similarity Searching Using 2D Fingerprints and Multiple Reference Compounds

“…In only a single instance (for TGD and activity class INO), SVM-"0.001%" produced a considerably lower recovery rate (8.5%) than the 1-NN, 5-NN, and centroid techniques (18.6-22.5%). However, five active and 14 database molecules represent a much smaller compound set than typically used for SVM training (see, for example, Wilton et al 13 ), and the high SVM performance level using this very small training set was not expected. In fact, Table 2 shows that differences in recovery rates of up to 30% were observed in favor of SVM-"0.001%" (e.g., for classes GPA, HIV using Molprint2D or IL1, THR using TGD).…”

“…When large training sets of 6000 active and up to 60 000 inactive molecules were used on the combined pesticide data sets, the machine learning methods, especially SVM, outperformed Unity 2D similarity rankings. 13 SVMs were originally developed for binary classification problems and have become popular in the chemoinformatics field. [16][17][18] In a typical SVM analysis, training compounds belonging to two different classes (e.g., active versus inactive) are projected into chemical reference space and a separating hyperplane is derived.…”

“…Here, we report the results of systematic search calculations on 10 different sets of pharmaceutically relevant compounds and five different types of fingerprints to compare the different ranking techniques. While SVMs are usually trained with fairly large compound sets, for example, see Wilton et al, 13 we used data sets for training comprising only five active molecules and between 14 and 14 423 randomly chosen database compounds. These compound sets were chosen in order to mimic practical virtual screening situations where often only a few active compounds are available.…”