-
Kernel learning for ligand-based virtual screening:discovery of a new PPARgamma agonist
(2010)
- Poster presentation at 5th German Conference on Cheminformatics: 23. CIC-Workshop Goslar, Germany. 8-10 November 2009 We demonstrate the theoretical and practical application of modern kernel-based machine learning methods to ligand-based virtual screening by successful prospective screening for novel agonists of the peroxisome proliferator-activated receptor gamma (PPARgamma) [1]. PPARgamma is a nuclear receptor involved in lipid and glucose metabolism, and related to type-2 diabetes and dyslipidemia. Applied methods included a graph kernel designed for molecular similarity analysis [2], kernel principle component analysis [3], multiple kernel learning [4], and, Gaussian process regression [5]. In the machine learning approach to ligand-based virtual screening, one uses the similarity principle [6] to identify potentially active compounds based on their similarity to known reference ligands. Kernel-based machine learning [7] uses the "kernel trick", a systematic approach to the derivation of non-linear versions of linear algorithms like separating hyperplanes and regression. Prerequisites for kernel learning are similarity measures with the mathematical property of positive semidefiniteness (kernels). The iterative similarity optimal assignment graph kernel (ISOAK) [2] is defined directly on the annotated structure graph, and was designed specifically for the comparison of small molecules. In our virtual screening study, its use improved results, e.g., in principle component analysis-based visualization and Gaussian process regression. Following a thorough retrospective validation using a data set of 176 published PPARgamma agonists [8], we screened a vendor library for novel agonists. Subsequent testing of 15 compounds in a cell-based transactivation assay [9] yielded four active compounds. The most interesting hit, a natural product derivative with cyclobutane scaffold, is a full selective PPARgamma agonist (EC50 = 10 ± 0.2 microM, inactive on PPARalpha and PPARbeta/delta at 10 microM). We demonstrate how the interplay of several modern kernel-based machine learning approaches can successfully improve ligand-based virtual screening results.
-
Virtual screening for PPAR-gamma ligands using the ISOAK molecular graph kernel and gaussian processes
(2009)
- For a virtual screening study, we introduce a combination of machine learning techniques, employing a graph kernel, Gaussian process regression and clustered cross-validation. The aim was to find ligands of peroxisome-proliferator activated receptor gamma (PPAR-y). The receptors in the PPAR family belong to the steroid-thyroid-retinoid superfamily of nuclear receptors and act as transcription factors. They play a role in the regulation of lipid and glucose metabolism in vertebrates and are linked to various human processes and diseases [1]. For this study, we used a dataset of 176 PPAR-y agonists published by Ruecker et al [2]. Gaussian process (GP) models can provide a confidence estimate for each individual prediction, thereby allowing to assess which compounds are inside of the model's domain of applicability. This feature is useful in virtual screening, where a large fraction of the tested compounds may be outside of the model's domain of applicability. In cheminformatics, GPs have been applied to different classification and regression tasks using either radial basis function or rational quadratic kernels based on vectorial descriptors [4,5]. We used a graph kernel based on iterative similarity and optimal assignments (ISOAK, [3]) for non-linear Bayesian regression with Gaussian process priors (GP regression, [4]). A number of kernel-based learning algorithms (including GPs) are capable of multiple kernel learning [5], which allows combining heterogeneous information by using multiple kernels at the same time. In this work, we combined rational quadratic kernels for vectorial molecular descriptors (MOE2D, CATS2D and Ghose-Crippen fragment descriptors) with the ISOAK graph kernel. We evaluated our methodology in different ranking and regression settings. Ranking performance was assessed using the number of false positives within the top k predicted compounds. Predicted compounds were ranked based on both predicted binding affinity and the confidence in each prediction. In the regression setting, we employed standard loss functions like mean absolute error (MEA) and root mean squared error. The established linear ridge regression (LRR) and support vector regression (SVR) algorithms served as baseline methods. In addition to standard test/training splits and cross-validation, we used a clustered cross-validation strategy where clusters of compounds are left out when constructing training sets. This results in less optimistic results, but has the advantage of favouring more robust and potentially extrapolation-capable algorithms than standard training/test splits and normal cross-validation. In the regression setting, both GP and SVR models performed well, yielding MAEs as low as 0.66 +- 0.08 log units (clustered CV) and 0.51 +- 0.3 log units (normal CV). In the ranking setting, GPs slightly outperform SVR (0.21 +- 0.09 log units vs. 0.3 +- 0.08 log units). In conclusion, Gaussian process regression using simultaneously – via multiple kernel learning – the ISOAK molecular graph kernel and the rational quadratic kernel (with standard molecular descriptors) performs excellent in retrospective evaluation. A prospective evaluation study is currently in progress.
