(1049-B) AI-driven method for identification of hits from phenotypic screening with Cell Painting Assay.

Abstract: High Content Screening is a powerful technology commonly used in phenotypic drug discovery. Currently, development of high content screening assays involves multiple technical processes, including cycles of wet-lab experiments and image/ data analysis. This process can be streamlined through the application of Cell Painting, an unbiased, target agnostic high content screening assay. The method, based on simple fluorescent staining of the 8 most important cellular compartments, generates imaging data that can be applied to identify biologically active compounds against various molecular targets without the need for development of different assays.
The key element in the process of identification of new active small molecules using Cell Painting Assay is the analysis of the results. Cell Painting generates multiparametric data from thousands of morphological features which are extracted from acquired microscopic images. Analysis of such complex datasets remains the biggest challenge that hinders researchers from using this approach for the hit identification purpose.
To tackle the problem, Ardigen developed a Deep Learning hit identification method as a part of its phenAID platform. The method is trained to predict compounds that induce the most similar phenotype to the reference compound, using images as a source of truth. The resulting output of the model is a ranking of possible hits with their probabilities.
To test this approach, we have used a JUMP-CP data set containing HCS images of ~120 k compounds created by 10 different laboratories. For each compound, multiple replicates were acquired across different partner sites, in each sample plate a negative control (DMSO) and a set of 8 positive controls with known mechanisms of action were included. The positive controls were selected to ensure diversity of targets and phenotypes. Controls were used as the reference points for identifying hits in datasets coming from different sources, and to test the robustness of our method comparing hits across sources. Additionally, we explored the chemical diversity of predicted hits, to ensure the method's viability in the drug design process.