Learning Single-Cell Distances from Cytometry Data.

Area of Science:

• Cytometry
• Computational Biology
• Data Science

Background:

• Automated cell population identification in cytometry relies heavily on distance metrics.
• The Euclidean distance is commonly used but may not be optimal for complex datasets.
• Single-cell labels offer an opportunity to derive more informative distance metrics.

Purpose of the Study:

• To develop and validate an optimal Mahalanobis distance metric for cell population identification using single-cell labels.
• To demonstrate the superiority of the Mahalanobis distance over the Euclidean distance in cytometry data analysis.
• To show the applicability of this method across different cytometry techniques and cell types.

Main Methods:

• Exploiting single-cell labels to derive a data-driven Mahalanobis distance metric.
• Applying the derived metric to flow cytometry data of microbial cells.
• Applying the derived metric to mass cytometry data of human blood cells.
• Evaluating performance with clustering methods.

Main Results:

• The Mahalanobis distance metric significantly improved cell population identification compared to the Euclidean distance.
• The method demonstrated effectiveness on diverse cytometry datasets (microbial flow cytometry, human blood mass cytometry).
• Improved cell identification was observed when employing clustering methods with the Mahalanobis distance.

Conclusions:

• An optimal Mahalanobis distance metric derived from data enhances automated cell identification in cytometry.
• This approach offers improved accuracy and can be applied to multiple samples under consistent experimental conditions.
• Utilizing advanced distance metrics is crucial for advancing the performance of cytometry data analysis techniques.