Modeling of Compound Profiling Experiments Using Support Vector Machines

Abstract

Profiling of compounds against target families has become an important approach in pharmaceutical research for the identification of hits and analysis of selectivity and promiscuity patterns. We report on modeling of profiling experiments involving 429 potential inhibitors and a panel of 24 different kinases using support vector machine (SVM) techniques and naïve Bayesian classification. The experimental matrix contained many different activity profiles. SVM predictions achieved overall high accuracy due to consistently low false-positive and consistently high true-negative rates. However, predictions for promiscuous inhibitors were affected by false-negative rates. Combined target-based SVM classifiers reached or exceeded the performance of SVM profile prediction methods and were superior to Bayesian classification. The classifiers displayed different prediction characteristics including diverse combinations of false-positive and true-negative rates. Predicted and experimentally observed compound activity profiles were compared in detail, revealing activity patterns modeled with different accuracy.

QSAR-derived affinity fingerprints (part 1): fingerprint construction and modeling performance for similarity searching, bioactivity classification and scaffold hopping

“…For compound biological activity prediction, various Quantitative Structure-Activity Relationship (QSAR) methods have been developed [42][43][44]. Recently, several groups adopted QSAR models to predict compound activity across the human kinome and generated corresponding affinity fingerprints [45][46][47][48]. In the Profile-QSAR (pQSAR) method [45], Naïve Bayes models were trained on 115 Novartis proprietary kinase assays.…”

Compound Structure‐Independent Activity Prediction in High‐Dimensional Target Space

“…[14] Given the attractiveness of these applications, it is not surprising that a number of computational investigations have attempted multi-target activity predictions. [15][16][17][18][19][20] From a methodological point of view, experimental profiling data can also serve as an input to compound activity prediction. Instead of predicting compound-target interactions based on chemical structures, which represents a conventional approach in chemoinformatics, bioactivitybased descriptors of compounds have been increasingly used in the past years.…”

“…[19,20] In a recent study, it was attempted to reproduce a kinase profiling experiment involving a set of 429 pyridinyl-imidazole ATP site-directed inhibitors tested against a panel of 24 different kinases. [20] Because the design of these kinase inhibitors was primarily focused on the p38-a (MAPK14) kinase, a popular cancer target, the compounds were often inactive against distantly related kinases. Hence, active data points were sparsely distributed over the profiling matrix.…”

“…Hence, active data points were sparsely distributed over the profiling matrix. [20] Nonetheless, a significant variety of compound profiles was observed. Under these conditions, serially applied target-based support vector machine (SVM) classifiers using structural fingerprint descriptors for compounds predicted profiles more accurately than profilebased regression or structural SVM models and naïve Bayesian (NB) classifiers.…”

“…Under these conditions, serially applied target-based support vector machine (SVM) classifiers using structural fingerprint descriptors for compounds predicted profiles more accurately than profilebased regression or structural SVM models and naïve Bayesian (NB) classifiers. [20] In this study, we have addressed a principally different activity prediction task using a kinase profiling data set consisting of 72 reference inhibitors including marketed drugs and 383 kinases covering major parts of the kinome. In this case, profiling data were available for only a limited number of compounds but many kinases, thus representing a high-dimensional target space with only little compound coverage.…”