UC Santa Cruz

Natural products have been a cornerstone in drug discovery, contributing to over 1000therapeutics in the past three decades. Traditional methods, such as the "grind and find" approach, have evolved with technological advancements, enabling high-throughput and high-content screening of chemical compounds. These advanced screening methods allow for measuring complex phenotypes, particularly through image-based assays that capture subtle cellular morphological changes. However, the complexity of the resulting data necessitates robust algorithms and workflows to accurately profile and interpret these phenotypes. First, this thesis presents the refactorization of the HistDiff algorithm used in the Cytological Profiling data processing workflow, tailored for high-content image-based phenotypic screening. The HistDiff chapter focuses on the improvements made to the algorithm, transitioning it from JAVA/Groovy to Python and enhancing its functionality and performance. This chapter also discusses the application of these enhancements to specific ongoing projects. Next, I introduce a classification methodology inspired by the BLAST sequence alignment tool. This method, designed to hypothesize the mechanism of action (MOA) based on pairwise associations of phenotypic fingerprints, has transitioned into a formal machine learning classifier, and aims to provide an accurate and accessible prediction framework. Finally, I outline the development of a web application to serve as the backbone of an ecosystem that integrates raw output from phenotypic screens with processed summaries, allowing users to interact with data efficiently. This application aims to be a powerful tool for researchers, enabling the exploration and interpretation of complex phenotypic screening data and ultimately facilitating the discovery of new therapeutic compounds.