Abstract
High Content Screening (HCS) platforms allow screening living cells under a wide range of experimental conditions and give access to a whole panel of cellular responses to a specific treatment. The outcome is a series of cell population images. Within these images, the heterogeneity of cellular response to the same treatment leads to a whole range of observed values for the recorded cellular features. Consequently, it is difficult to compare and interpret experiments. Moreover, the definition of phenotypic classes at a cell population level remains an open question, although this would ease experiments analyses. In the present work, we tackle these two questions. The input of the method is a series of cell population images for which segmentation and cellular phenotype classification has already been performed. We propose a probabilistic model to represent and later compare cell populations. The model is able to fully exploit the HCS-specific information: "dependence structure of population descriptors" and "within-population variability".
Introduction
High Content Screening (HCS) produces massive amounts of imaging data, allowing researchers to observe how thousands of different treatments affect cell populations. Traditionally, researchers analyze this by identifying individual cell phenotypes and then averaging those results across a whole well. However, this "averaging" approach ignores two critical pieces of information: within-population variability and the dependence structure between cellular descriptors. Cells in the same well don't all behave identically, and cellular descriptors are often correlated (e.g., cell density affecting growth phase proportions). The authors aim to bridge the gap between single-cell data and population-level interpretation by creating a mathematical framework that treats a cell population as a structured entity rather than just a mean value.
Methods
The authors propose a hierarchical mixture model to perform unsupervised population phenotyping. The process follows a multi-step pipeline from raw images to population descriptors.
The model assumes each well belongs to a latent "population phenotype" class. Within a well, multiple images (fields) are treated as samples from a multivariate distribution associated with that class. To model the "dependence structure," the authors use Gaussian copulas. Unlike standard Gaussian models that struggle with bounded data (like proportions), copulas allow researchers to model the marginal distributions (e.g., Gamma for cell counts, Beta for proportions) separately from the correlation structure between them. An Expectation-Maximization (EM) framework is used to learn the parameters and assign wells to classes.
Key Results
The model was validated using a real-world siRNA screening dataset on a shA673-1C Ewing sarcoma-derived cell line, encompassing 2688 images from 672 wells. The copula-based hierarchical model consistently achieved higher log-likelihood on test data compared to both the Gaussian mixture and the baseline models across various numbers of phenotypic classes.
Beyond model fit, the framework successfully identified experimental outliers, such as imaging artifacts or poor staining, by isolating them into distinct, small classes. It also effectively separated positive (KIF11) and negative (GL2) controls in a multivariate space, demonstrating its power for novelty detection in siRNA screens.
Conclusion
While the copula-based model provides a richer analysis by accounting for biological dependencies and variability, it is computationally more intensive than Gaussian versions. Future work includes integrating single-cell measurements directly into the model and further biological validation of the inferred phenotypic classes to determine their functional significance.
Summary updated