Multimodal Approach to MoA Prediction Based on Cell Painting Imaging and Chemical Structure Data
Cell Painting protocol is currently emerging as the go-to method for phenotypic screening in drug discovery. This approach yields massive amounts of information encoded in multicolor
images at single cell resolution, which in turn raises issues with proper analysis. In order to address this problem a range of proprietary and open source algorithms for cellular feature extraction has been developed, allowing for a semi-automatic analysis of high content data with decent reliability.
However, currently available solutions miss out on a significant portion of data available in phenotypic screening experiments: the chemical structure of tested compounds. While deriving properties and predictions from compounds’ structures is a wholly different branch of cheminformatics, there were few efforts to combine the information from images and chemical structures into a single, more robust data structure.
We explored this approach using human-defined descriptors (CellProfiler for images and ECFP for compounds) acquiring improved results compared to each of the respective methods alone. To improve even further, we used deep representations of images (GapNet) and compounds (R-MAT), achieving state-of-the-art results for mode of action prediction in high content screening results analysis.