Chemogenomics: Models of Protein-Ligand Interaction Space

Abstract: The large majority of the currently used drugs are small molecules that interact with proteins. Understanding protein-ligand recognition is thus central to drug discovery and design. Improved experimental techniques have resulted in an immense growth of drug target information. This has stimulated the development of chemogenomics and proteochemometrics (PCM) that take target information as well as ligand information into account to study the genomic effect of potential drugs.This thesis is concerned with modeling protein-ligand recognition, and the aim is to develop models that generalize to the entire protein-ligand space. To this end, protein-ligand interaction data has been extracted and manually curated from public databases, protein and ligand descriptors have been computed, and predictive models have been induced with machine-learning methods.An introduction to chemogenomics, machine learning, and PCM modeling is given in the thesis summary, which is followed by five research papers. Paper I shows that it is possible to induce interpretable models with a non-linear rule-based method, and paper II demonstrates that local descriptors of protein structure may be used to induce PCM models that cover proteins differing in sequence and fold. In paper III, such local descriptors are used to induce a PCM model on a large dataset that includes all major enzyme classes. This demonstrates that the local descriptors may be used to induce generalized models that span the entire known structural enzyme-ligand space. Paper IV describes a step towards proteome-wide PCM models, and shows that it is possible to predict high- and low-affinity complexes using a set of protein and ligand descriptors that do not require knowledge of 3D structure. Finally, paper V presents a method to visualize and compare protein-ligand chemogenomic subspaces, which may be used to predict unwanted cross-interactions of drugs with other proteins in the proteome.