Merging Mixture Components for Cell Population Identification in Flow Cytometry

Greg Finak, Ali Bashashati, Ryan Remy Brinkman, Raphael GottardoAdv. Bioinformatics 2009 · OpenReview

TL;DR

This paper introduces "flowMerge," a novel framework for automatically identifying cell populations in complex flow cytometry data. By merging over-segmented clusters, flowMerge improves model fit and provides a more accurate count of distinct cell subpopulations compared to existing methods, making high-throughput analysis more reliable.

What Problem Does This Paper Address?

Flow cytometry (FCM) is a powerful technique for analyzing millions of cells per day, but its data analysis, particularly "gating" (identifying cell populations), is often manual and subjective. This manual process introduces variability and limits the potential of high-throughput FCM. While automated methods exist, they struggle with the complex, asymmetric, and sometimes overlapping shapes of cell populations commonly found in FCM data. Existing model-based approaches, like Gaussian Mixture Models (GMMs) and the flowClust method, often either over-segment populations (requiring many components for a single population) or underestimate the true number of distinct populations, leading to inaccurate analyses. The core challenge is to find an automated method that accurately identifies distinct cell populations while handling complex data distributions and providing a parsimonious representation.

Methods

The authors extend the "cluster merging" algorithm, originally proposed for GMMs, to work with the more flexible multivariate t-mixture models used in the flowClust framework. Here's a breakdown of their approach:

  • flowClust Framework: They leverage flowClust, which fits multivariate t-distributions to FCM data. This model allows for data transformations (Box-Cox) to make distributions more symmetric and uses t-distributions to better handle outliers and asymmetric populations.
  • Cluster Merging Algorithm: The core innovation is an iterative merging process. It starts with an initial clustering (e.g., from flowClustBIC) and progressively merges pairs of overlapping clusters that minimize the overall clustering entropy. This process aims to reduce over-segmentation and create more biologically meaningful populations.
  • Automated Number of Clusters Selection: Instead of relying solely on criteria like BIC or ICL (which have limitations for FCM data), they use changepoint analysis on the plot of clustering entropy versus the number of clusters. This helps identify the point where remaining clusters are well-separated, thus determining the optimal number of merged clusters.
  • Summarizing Merged Components: To facilitate downstream analysis, they developed a method to represent each merged cluster with a single set of parameters (mean and covariance) by matching the moments of the merged distribution. This provides a concise parametric representation of complex populations.
  • Borderline Case Identification: They also implemented criteria to flag samples where the merging algorithm might fail, allowing for manual inspection.

Datasets

  • Real-world Data: The primary dataset used for evaluation is from 137 samples of Chronic Lymphocytic Leukemia (CLL) patients, provided by the BC Cancer Agency. This dataset involves multiple antibody markers and samples from 18 individuals.
  • Simulated Data: To rigorously test their method, they generated 100 simulated datasets. These simulations were based on the empirical distribution of a real CD8 vs. CD4 projection from a CLL sample, ensuring the simulated data mimicked realistic FCM complexities without introducing bias towards any specific model.

Key Insights and Ideas

  • Merging as a Solution to Over-segmentation: The key insight is that existing methods often over-segment complex populations. By merging these over-segmented clusters, their approach achieves a better balance between model fit and parsimony.
  • Entropy as a Measure of Population Separation: The paper effectively uses clustering entropy to guide the merging process. Lower entropy indicates better-separated clusters, and the algorithm iteratively reduces entropy by merging overlapping components.
  • Changepoint Analysis for Optimal Cluster Count: Applying changepoint detection to the entropy-number of clusters plot is a clever way to automate the selection of the "correct" number of cell populations, overcoming limitations of traditional model selection criteria.
  • Parametric Summarization of Complex Populations: The ability to represent complex, merged populations with a single set of parameters is crucial for integrating automated gating into existing analysis pipelines and enabling downstream statistical analysis.
  • Identifying Problematic Marker Combinations: The analysis of normalized entropy across different antibody combinations revealed specific marker sets (e.g., CD10/CD11c/CD20) that consistently presented challenges for automated gating, highlighting areas for future methodological improvements.

Key Results

  • Improved Model Fit and Cluster Count: The flowMerge framework consistently provided a better fit to complex FCM data and a more accurate estimate of the number of distinct cell subpopulations compared to flowClustBIC, flowClustICL, and GMMs.
  • CLL Data Analysis: On the CLL dataset, flowMerge identified a median of 4 populations (range 2-8), generally falling between the over-segmenting flowClustBIC (median 6) and under-segmenting flowClustICL (median 2).
  • Simulation Study: In simulations, flowMerge achieved the lowest misclassification rate (4.2%) when compared to flowClust and GMMs, especially when identifying rare cell subpopulations. While flowMerge sometimes overestimated the true number of clusters, these extra clusters primarily served to model outlier populations.
  • Robustness: The method demonstrated robustness in identifying key populations like lymphocytes, even when they were represented by multiple components in other models.

Limitations and Future Work

  • Borderline Cases: The authors acknowledge that the merging algorithm can fail in certain scenarios, particularly with highly overlapping or complex distributions. They have implemented criteria to flag these cases for manual inspection.
  • Overestimation of Clusters: While flowMerge generally provides a better estimate, it can still overestimate the number of clusters, though these extra clusters are often associated with outlier populations.
  • Future Work: The paper suggests that further investigation into problematic marker combinations and the development of strategies to improve automated gating for these cases would be valuable. The framework's ability to handle repeated-measures designs in FCM data is also an area for future exploration.

Summary updated

Abstract

We present a framework for the identification of cell subpopulations in flow cytometry data based on merging mixture components using the flowClust methodology. We show that the cluster merging algorithm under our framework improves model fit and provides a better estimate of the number of distinct cell subpopulations than either Gaussian mixture models or flowClust, especially for complicated flow cytometry data distributions. Our framework allows the automated selection of the number of distinct cell subpopulations and we are able to identify cases where the algorithm fails, thus making it suitable for application in a high throughput FCM analysis pipeline. Furthermore, we demonstrate a method for summarizing complex merged cell subpopulations in a simple manner that integrates with the existing flowClust framework and enables downstream data analysis. We demonstrate the performance of our framework on simulated and real FCM data. The software is available in the flowMerge package through the Bioconductor project.
BibTeX
@inproceedings{finak2009merging,
  author = {Greg Finak and Ali Bashashati and Ryan Remy Brinkman and Raphael Gottardo},
  title = {{Merging Mixture Components for Cell Population Identification in Flow Cytometry}},
  booktitle = {Adv. Bioinformatics},
  year = {2009},
  url = {https://openreview.net/forum?id=D6om-M5ta0H},
}
Figure