Robust High-Throughput Imaging Analysis with Wasserstein Geodesic Transformations

Abstract

High-throughput cell imaging has been increasingly used in drug discovery to simultaneously profile the morphological response of cells to thousands of compounds using high-resolution microscopy and automated image analysis. Such experiments characterize thousands of image features in millions of cells across thousands of experimental conditions. Analytical difficulties arise with this scale of analysis as many features have distributions with extremely long tails, high skewness, remote outliers, and high-leverage points. This makes important signals difficult to find and means analyses are often sensitive to individual observations or features. This work considers a recent high-quality Cell Painting dataset profiling compounds from the EU-OPENSCREEN consortium. The study perturbs HepG2 human liver cancer cells in order to morphologically profile cellular response to the compounds. Without adjustment, analysis of the imaging data is hampered by long-tailed distributions and outliers. To combat this, we introduce Wasserstein Geodesic Transformations (WGTs), a new approach that adaptively moves features in Wasserstein space to make downstream analysis less ad-hoc, more stable, and more scalable. In application to the Cell Painting data, WGTs substantially improve data analysis by enhancing visualization, improving compound clustering, and stabilizing analyses. They also help uncover unwanted spatial effects arising from plate layout, explaining some outlying compound responses. More broadly, the adaptivity of WGT makes it a promising tool a wide-range of cell imaging pipelines.