Kernel methods in chemo- and bioinformatics

Meer over het boek

This thesis explores solutions to machine learning problems in chemo- and bioinformatics using kernel methods, a family of algorithms known for their solid theoretical foundation and practical success. It begins with a review of the fundamentals of kernel machines, followed by a novel algorithm for model selection in Support Vector Machines (SVMs) for classification and regression. This algorithm, based on global optimization theory, is efficient and does not rely on specific properties of the kernel function. Experimental results show promising improvements over existing methods. The focus then shifts to applications in drug discovery, particularly the ADME in silico prediction problem. The thesis investigates descriptor and graph-based representations of molecules, proposing a descriptor selection algorithm that enhances statistical stability. A new class of specialized kernel functions is introduced for comparing molecules at a graph-based level, integrating expert knowledge and various molecular similarity notions into a unified SVM model. Additionally, a reduced graph representation for molecular structures is proposed, condensing certain elements into single nodes, leading to improved prediction performance while maintaining computational efficiency. The thesis also examines the features of membrane potential (MP) that influence action potentials (APs) in cortical neurons. SVMs are trained to predict AP occurrence usin