Classifying flow cytometry data using Bayesian analysis helps to distinguish ALS patients from healthy controls

Saskia Räuber, Christopher Nelke, Christina B. Schroeter, Sumanta Barman, Marc Pawlitzki, Jens Ingwersen, Katja Akgün, René Günther, Alejandra P. Garza, Michaela Marggraf, Ildikò Rita Dunay, Stefanie Schreiber, Stefan Vielhaber, Tjalf Ziemssen, Nico Melzer, Tobias Ruck, Sven G. Meuth, Michaël HertyFrontiers in Immunology 2023 · DOI

TL;DR

Researchers have developed a novel diagnostic approach for Amyotrophic Lateral Sclerosis (ALS) by applying Bayesian network analysis to raw flow cytometry data. By building a "Healthy Control tree" to define normal immune signatures, the algorithm can identify ALS patients with near-perfect accuracy (AUC up to 1.0) without the need for traditional, data-stripping dimensionality reduction.

What Problem Does This Paper Address?

Multidimensional flow cytometry (mFC) is a cornerstone of immunology, allowing scientists to look at dozens of markers on millions of individual cells. However, analyzing this "big data" is a bottleneck. Traditional manual gating is slow and prone to human bias.

To solve this, many researchers turn to automated computational tools. The catch? Most of these tools rely on dimensionality reduction—essentially "squashing" the data to make it manageable. This process often discards rare cell populations or subtle marker dependencies that might contain the actual "signal" of a disease. The authors sought an unbiased, automated method that retains the full complexity of the raw data to distinguish ALS patients from healthy individuals.

Methods

The authors moved away from traditional clustering and instead used Bayesian Network Analysis.

  1. The HC Tree: They used the Chow-Liu algorithm to create a mathematical model (a dependency tree) of how different markers relate to one another in a "Healthy Control" (HC) cohort.
  2. Probability Scoring: When a new sample (either ALS or HC) is fed into the model, the algorithm calculates the probability ($q$) of each individual cell belonging to that healthy distribution.
  3. The "NoZ" Metric: The key classifier is the Number of Zeros (NoZ)—the relative number of cells in a sample where the probability $q$ is exactly zero. Essentially, it counts how many cells in a patient's blood are "impossible" or highly unlikely according to the healthy model.
  4. Comparison: They compared their results against Citrus, a popular commercially available algorithm for automated flow cytometry analysis.

Datasets

The study utilized two independent cohorts from German medical centers:

  • Cohort I (Dresden): 68 ALS patients and 35 Healthy Controls. This cohort used a 14-marker panel (including CD3, CD4, CD8, CD38, and CD25).
  • Cohort II (Magdeburg): 55 ALS patients and 30 Healthy Controls. This served as a validation set and used a different marker panel (including CD14, CD16, and HLA-DR) to test the algorithm's flexibility.

Key Insights and Ideas

  • Embracing the Raw Data: The most significant "aha moment" is that the algorithm performs better on raw, ungated data than on pre-processed data. While humans need to "clean" data to understand it, the Bayesian network thrives on the noise and complexity of the full dataset.
  • The "Outlier" Philosophy: Instead of trying to define exactly what an "ALS cell" looks like, the model defines what a "Healthy cell" looks like and then measures the "unhealthiness" of a sample.
  • Marker Synergy: The model revealed that certain combinations of markers (like CD4 and CD38) were more likely to yield "zero probability" cells in ALS patients, suggesting these specific immune pathways are significantly altered in the disease.
  • Algorithmic Superiority: When using raw data, the Bayesian approach outperformed Citrus. Citrus struggled with the high dimensionality of ungated files, whereas the Bayesian network's error rate remained significantly lower.

Key Results

  • High Diagnostic Accuracy: In the Dresden cohort, the algorithm achieved an AUC of 0.97, correctly identifying 64 out of 68 ALS cases.
  • Perfect Validation: In the independent Magdeburg cohort, the algorithm achieved a perfect AUC of 1.0, identifying every single ALS patient correctly.
  • Robustness: The model maintained high performance (AUC > 0.90) regardless of the number of markers or patients used to build the initial HC tree, proving it isn't overly sensitive to minor parameter changes.
  • Marker Identification: The study identified CD4, CD38, and CD20 as high-contributing markers for the "zero probability" classification, providing a starting point for further biological investigation into ALS immunology.

Limitations and Future Work

While the results are impressive, the authors note a few hurdles before this hits the clinic:

  • The Specificity Challenge: Currently, the model only distinguishes ALS from healthy controls. In a real clinical setting, it needs to distinguish ALS from "look-alike" neurological diseases (differential diagnosis).
  • Interpretability: While the model tells us that a sample is different, it’s still somewhat of a "black box" regarding the specific biological function of those outlier cells.
  • Confounding Factors: Future studies need to account for variables like patient age, sex, and medications, which could potentially shift the immune profile independently of the ALS diagnosis.

The authors conclude that Bayesian network analysis is a powerful, scalable tool that could eventually be integrated into routine clinical workups for complex neurodegenerative diseases.

Summary updated

Abstract

Bayesian network analysis might present a novel approach for classifying mFC data, which does not rely on reduction techniques, thus, allowing to retain information on the entire dataset. Future studies will have to assess the performance when discriminating clinically relevant differential diagnoses to evaluate the complementary diagnostic benefit of Bayesian network analysis to the clinical routine workup.
BibTeX
@inproceedings{rauber2023classifying,
  author = {Saskia Räuber and Christopher Nelke and Christina B. Schroeter and Sumanta Barman and Marc Pawlitzki and Jens Ingwersen and Katja Akgün and René Günther and Alejandra P. Garza and Michaela Marggraf and Ildikò Rita Dunay and Stefanie Schreiber and Stefan Vielhaber and Tjalf Ziemssen and Nico Melzer and Tobias Ruck and Sven G. Meuth and Michaël Herty},
  title = {{Classifying flow cytometry data using Bayesian analysis helps to distinguish ALS patients from healthy controls}},
  booktitle = {Frontiers in Immunology},
  year = {2023},
  doi = {10.3389/fimmu.2023.1198860},
}
Figure