(1218-C) Application of parallel machine learning methodologies to surface mechanistically distinct classes of compound hits in a high-content imaging screen for inflammasome inhibitors
Tuesday, February 6, 2024
12:00 PM – 1:00 PM EST
Location: Exhibit Halls AB
Abstract: AI/ML technologies are revolutionizing phenotypic screening approaches. Unlike prior phenotypic readout quantifications, these technologies offer many biologically interpretable analytical approaches. In this poster, we describe three complementary ML methods that were applied to the fluorescence imaging data of a single high throughput assay with the goal of surfacing mechanistically distinct compound classes.
Inflammasome therapeutic molecule discovery typically relies on in vitro activation assays with terminal pathway readouts that are honed to identify robust, late-pathway inhibitors. These assays cannot distinguish between compound mechanisms of action, which limits the optimal coupling of therapeutics to the vast array of inflammasome-related pathologies. We therefore conducted a high-content imaging screen of human PBMCs activated with two independent inflammasome stimuli and used AI-facilitated analysis tools to identify and functionally define inflammasome inhibitors. Specifically, we employed the following three approaches after first passing compounds through a toxicity score filter: (1) automated ASC speck quantification, which was used as a ground truth for inflammasome activation; (2) targeted scoring that uses differentially weighted ML-derived features associated with inflammasome activation and inhibition; (3) Compound mediated distribution shifts (CMDS) of morphological profile vector norms
823 non-toxic compounds were identified as hits using the targeted scoring method. 750 of these (91.1%) were cross-validated using the CMDS method, all compounds called hits (%) using these rubrics had measurable decreases in ASC specks under at least one of the two activation conditions tested. Altogether, there was strong agreement for hit designations between these methods, which was further improved when more stringent scoring thresholds were applied.
Identified compound hits were then categorized into functional classes using a set of scores that apply differential weight to biological and ML-derived features associated with early stages of activation, reversible pathway convergence points, or the terminal, pyroptotic step where inflammasome pathways ultimately converge. The distance measurement was similarly used to resolve early and late inflammasome pathway inhibitors by setting distance thresholds based on the distribution of the control conditions. 91.4% of compounds (192/210) identified as late-stage inhibitors using the terminal step scoring rubric were also identified as late inhibitors using the CMDS method, indicating very close alignment and cross-validation between the two methods.
Altogether, we utilized ML and advanced multi-dimensional data analysis to surface and categorize mechanistically distinct classes of compound modulators from a high-content imaging screen. These methods could be broadly applied to therapeutic compound screens or used to resolve complex, physiologically relevant biological pathways.