|
Drug discovery and development is a complex, long and expensive process involving searches in a multidimensional space and multi-objective optimization based on apparently unrelated target functions. The discovery of a new "small molecule" medicine depends on the identification of biologically active chemical entities ("hit" compounds), followed by chemical optimization of their biological activity (affinity and selectivity for their target), toxicity (safety), pharmacokinetics (how to reach the target tissue or organ), but also optimization of chemical and metabolic stability, administration modalities (formulation), among others. For several years, first at UC, and more recently at BSIM, we have dedicated our research effort to understand the molecular mechanisms and develop therapeutic strategies for a rare, fatal, neurodegenerative disease -Familial Amyloid Polyneuropathy (TTR-FAP). This pathology originates from the formation of cytotoxic aggregates (amyloid) of the protein transthyretin (TTR) and has clinical manifestations involving the peripheral nervous system (PNS), the central nervous system (CNS) and the eye. At BSIM, we have been successful at identifying small molecules able to very efficiently stabilize the native form of TTR and thus inhibit the process of amyloid formation. In this seminar, we will discuss the challenges of optimizing some of these compounds to cross the cornea and stabilize the pool of TTR in the eye, thus inhibiting the ocular manifestations of TTR-FAP. Through mining of scientific literature and patent data, and interactions with ophthalmology experts, we assembled a data set comprised of 70 compounds: 35 compounds capable of penetrating the cornea and reach back-of-the-eye compartments in therapeutic concentrations ("cornea +" class), and 35 compounds unable to cross the cornea ("cornea -" class). For each compound, 3107 molecular descriptors (MD) were computed. A total of 4006 machine learning models were assessed for their performance at predicting corneal permeability: 424 k-nearest neighbours (KNN), 1116 variants of Random Forest (RF) and 2628 Support Vector Machines (SVM). Our results show that an ensemble modeling strategy is an adequate approach for such a demanding problem, offering performances as high as 91% prediction accuracy.
|
