Comparison of Deep-Learning Models for Classification of Cellular Phenotype From Flow Cytometry Data

Area of Science:

• Computational Biology
• Immunology
• Machine Learning

Background:

• Flow cytometry is a powerful tool for analyzing cell populations.
• Automated analysis of high-dimensional flow cytometry data remains challenging.
• Deep learning offers potential for uncovering subtle cellular phenotypes.

Purpose of the Study:

• To compare the utility of deep learning models for automated phenotypic classification using flow cytometry data.
• To evaluate different deep neural network architectures for processing high-dimensional cytometry data.
• To assess the accuracy of deep learning in identifying subtle cellular phenotypes, such as sex differences in mice.

Main Methods:

• Three deep neural network architectures were evaluated: convolutional neural network (CNN) on raw data, multi-layer perceptron (MLP) on histograms, and MLP on hypervoxels.
• The models were trained and tested on peripheral blood cell populations from 2300 mouse samples.
• Classification accuracy was measured for a mutant phenotype identification task.

Main Results:

• A histogram-based multi-layer perceptron achieved the highest classification accuracy of 91%.
• Deep learning models demonstrated high sensitivity in distinguishing male from female mice based on blood cell attributes.
• The study successfully identified subtle cellular phenotypes not easily detected by visual analysis.

Conclusions:

• Deep learning models are effective for automated analysis of complex flow cytometry data.
• Histogram-based MLPs show significant promise for phenotypic classification in high-dimensional datasets.
• This approach has the potential to advance the detection of biologically meaningful phenotypes from cytometry data.