Abstract

Predictive Models for Mechanism of Action Classification from Phenotypic Assay Data

Alexander Baumann, Senior Director, DiscoveRx Corporation

We have previously described a platform of human cell based assays (BioMAP® systems) for compound characterization and phenotypic drug discovery. These assays contain a diversity of early passage primary human cell types and co-cultures and model complex tissue and disease states. Agents are tested in these models and compound activity profiles are generated by measuring changes in the levels of a set (8-12) of biomarkers (proteins, mediators, etc.) in each BioMAP system. Compounds and materials profiled in these systems include approved and failed drugs, small molecules from various target-based and phenotypic drug discovery programs, compounds from diversity and focused libraries, natural products collections and biologics, as well as environmental chemicals and nanomaterials. BioMAP profile signatures have been shown to distinguish compounds based on the mechanism of action (MoA) and target selectivity, and BioMAP activities have been correlated to in vivo biology. We have previously described using results from an unsupervised search of a large database of reference compound profiles to generate MoA hypotheses. Since this method is cumbersome and requires expert interpretation, however, a more automated and quantitative approach, such as that provided by machine learning is attractive to consider. We have now developed a useful method for assigning mechanism class to compounds and bioactive agents using machine learning. We have developed predictive models for 28 mechanism classes. These mechanism classes were selected to encompass safety and efficacy-related mechanisms, include both target-specific and pathway-based classes, and cover the most common mechanisms identified in phenotypic screens, such as inhibitors of mitochondrial and microtubule function, histone deacetylase and cAMP elevators. By providing quantified membership in specific mechanism classes, this approach is suitable for identification of off-target toxicity mechanisms as well as enabli