Comparison of Support Vector Machine and Artificial Neural Network Systems for Drug/Nondrug Classification

Abstract

Support vector machine (SVM) and artificial neural network (ANN) systems were applied to a drug/nondrug classification problem as an example of binary decision problems in early-phase virtual compound filtering and screening. The results indicate that solutions obtained by SVM training seem to be more robust with a smaller standard error compared to ANN training. Generally, the SVM classifier yielded slightly higher prediction accuracy than ANN, irrespective of the type of descriptors used for molecule encoding, the size of the training data sets, and the algorithm employed for neural network training. The performance was compared using various different descriptor sets and descriptor combinations based on the 120 standard Ghose-Crippen fragment descriptors, a wide range of 180 different properties and physicochemical descriptors from the Molecular Operating Environment (MOE) package, and 225 topological pharmacophore (CATS) descriptors. For the complete set of 525 descriptors cross-validated classification by SVM yielded 82% correct predictions (Matthews cc = 0.63), whereas ANN reached 80% correct predictions (Matthews cc = 0.58). Although SVM outperformed the ANN classifiers with regard to overall prediction accuracy, both methods were shown to complement each other, as the sets of true positives, false positives (overprediction), true negatives, and false negatives (underprediction) produced by the two classifiers were not identical. The theory of SVM and ANN training is briefly reviewed.

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

“…Our results are significantly better than those reported in ref 12. We show that (a) using the same amount of training data and after appropriate preprocessing the error rate can be reduced to 18.1% and (b) using more training examples we achieve a surprisingly low error rate of 10.2%.…”

“…For this more realistic data set we show in simulations that the best method achieves error rates at around 7% which we additionally confirmed in a blind-test experiment. We therefore were able to reduce the error rate for the task of classifying drugs (WDI) vs nondrugs (ACD) by more than 60% compared to Byvatov et al 12 and Sadowski and Kubinyi. 14 …”

“…We investigated methods that range from conventional decision trees, 15 Fishers Linear Discriminant 16 and Nearest Neighbor methods 17 to advanced learning techniques such as SVMs 5,7,8 and linear programming machines. 18,19 Contrary to Byvatov et al 12 we solely base our study on a Ghose-Crippen parametrization of the chemicals. 13 However, before feeding it into any of the learning machines we preprocess the GC descriptors by adding one and then computing the log.…”

“…Recent work by Byvatov et al 12 considered a large benchmark data set containing atomic descriptors of compounds that are either drugs (WDI) or nondrugs (ACD). Byvatov et al proposed a Support Vector Machine (SVM) based method that correctly classifies 80% (error rate 20%) of the compounds based on Ghose-Crippen (GC) descriptors (counting occurrences of atoms according to the 120 types defined in ref 13).…”

SVM-Based Feature Selection for Characterization of Focused Compound Collections

“…The approach complements related work on "drug-likeness" prediction 12 and extends it to target-and target-family specific sets of inhibitors.…”