Classifying ‘Drug-likeness' with Kernel-Based Learning Methods

Abstract

In this article we report about a successful application of modern machine learning technology, namely Support Vector Machines, to the problem of assessing the 'drug-likeness' of a chemical from a given set of descriptors of the substance. We were able to drastically improve the recent result by Byvatov et al. (2003) on this task and achieved an error rate of about 7% on unseen compounds using Support Vector Machines. We see a very high potential of such machine learning techniques for a variety of computational chemistry problems that occur in the drug discovery and drug design process.

Structure–Activity Relationships in Chromatography: Retention Prediction of Oligonucleotides with Support Vector Regression

“…In addition to the advantage of leading exactly to a globally optimal solution (unlike ANNs), support vector approaches have also proved themselves for practical use with chemical problems. [9,10] A model based on the determination and addition of the retention contributions of the nucleotides has been developed to predict the retention of oligonucleotides in ion-pair reversed-phase chromatography (IP-RPC). [11] This model provided satisfactory results at relatively high separating temperatures (60 8C) for cases in which secondary structures are less pronounced, whereas at lower temperatures, the influence of hairpin or partial double strands led to a poor correlation between the prediction and the experiment (own measurement results).…”

Gradual in Silico Filtering for Druglike Substances

“…Although support vector machines are known to yield very good results in performing classification tasks, the obtained accuracies are not throughout better than those of the decision trees as observed in a related study but rather similar. 20 Müller and co-workers compared a number of machine learning methods but used nonlinear kernel func- 20 Recently, Li et al applied a probabilistic support vector machine with a radial basis that yielded better results than with a linear kernel, as expected. 17 While the linear function corresponds to a separating line in a 2-dimensional descriptor space, radial basis and polynomial functions introduce curvatures.…”

“…7,[11][12][13][14][15][16][17] However, even support vector machines that are known to be among the most accurate methods for classification tasks have so far not provided fully satisfactory results. [17][18][19][20] This can be, at least in part, explained by the implicit difficulties that arise from the data set.…”

Using Machine Learning Methods to Predict Experimental High Throughput Screening Data

“…Using the majority vote for all tested techniques, the latter authors concluded that NN provided the best prediction of experimental results, followed by SVM. On the other hand, Müller et al [7] described an application of SVM to the problem of assessing the "drug-likeness" of a compound based on a given set of molecular descriptors. The authors concluded that in the drug-likeness analysis a polynomial SVM with a high polynomial degree (d = 11) allows for a very complex decision surface which could be used for prediction.…”