A statistical framework for high-content phenotypic profiling using cellular feature distributions

Abstract

High-content screening (HCS) uses microscopy images to generate phenotypic profiles of cell morphological data in high-dimensional feature space. While HCS provides detailed cytological information at single-cell resolution, these complex datasets are usually aggregated into summary statistics that do not leverage patterns of biological variability within cell populations. Here we present a broad-spectrum HCS analysis system that measures image-based cell features from 10 cellular compartments across multiple assay panels. We introduce quality control measures and statistical strategies to streamline and harmonize the data analysis workflow, including positional and plate effect detection, biological replicates analysis and feature reduction. We also demonstrate that the Wasserstein distance metric is superior over other measures to detect differences between cell feature distributions. With this workflow, we define per-dose phenotypic fingerprints for 65 mechanistically diverse compounds, provide phenotypic path visualizations for each compound and classify compounds into different activity groups.

SPACe (Swift Phenotypic Analysis of Cells): an open-source, single cell analysis of Cell Painting data

“…A version of our previously published (19) quality control (QC) pipeline was added in SPACe (Step 4); this step leverages the analysis of the empirical probability distribution of single cell features in control samples ( i.e., DMSO treated) to identify and discard outlier wells and to generate a reference distribution for each feature, defined as the median empirical distribution of all remaining DMSO wells. This reference distribution is then used to quantify the effect of perturbations based on the Earth Mover’s Distance (EMD), a metric that quantifies the dissimilarity between probability distributions and has been shown to work well for phenotypic screens (18). Additional details, including data normalization, can be found in the Methods section.…”

“…Nonetheless, while CP approaches have entered the mainstream for phenotypic screening, there is still a very active research effort to enhance robustness, processing speed, and sensitivity to best capture cell population heterogeneity. As AI-driven strategies have become more prevalent for high-dimensional and large-scale data analysis, integration of such strategies into CP and its variants has stimulated a major interest due to the promise of higher accuracy, faster processing times and the potential for fusing multimodal data (18,19).…”

“…Information extraction from CP images involves two main steps: firstly, identifying regions of interest (e.g., cellular substructures) and extracting relevant features, typically done using open-source ( i.e., CellProfiler, (13,14)) or commercial software; secondly, performing feature reduction and representation (1,15–17) to facilitate further downstream analyses such as clustering and classification. Due to its biological and analytical relevance, single cell data in HT imaging-based campaigns is now widely used both for data quality control and for hit identification associated with various treatments based upon clear phenotypic differences (6,18–23). Nonetheless, while CP approaches have entered the mainstream for phenotypic screening, there is still a very active research effort to enhance robustness, processing speed, and sensitivity to best capture cell population heterogeneity.…”

“…Nonetheless, while CP approaches have entered the mainstream for phenotypic screening, there is still a very active research effort to enhance robustness, processing speed, and sensitivity to best capture cell population heterogeneity. As AI-driven strategies have become more prevalent for high-dimensional and large-scale data analysis, integration of such strategies into CP and its variants has stimulated a major interest due to the promise of higher accuracy, faster processing times and the potential for fusing multimodal data (18,19). However, with the continual advancement of HT microscopes and laboratory automation, and resultant large datasets, a key roadblock appears to be how to analyze data efficiently, in a timely manner, and the availability of massive computational resources needed to carry out such analysis.…”

“…We recall that most of existing cell screening analytical platforms, while extracting information from individual cells, ultimately utilize only per well or per treatment central tendency values like mean or median for downstream analysis. A notable exception is the work from Pearson et al, (18) that demonstrated the potential advantage of interrogating single cell information by analyzing the statistical distribution of data within cellular populations. While the traditional per-well average approach has proven to be successful and sufficient for hit calling from single end point assays in large scale screening campaigns, it ignores the inherent phenotypic heterogeneity in a cell population and the fact that many biological responses do not follow a normal distribution (6,19,23).…”