Machine Learning Empowers Mass Cytometry: A New Breakthrough in Precision Analysis of Chronic Lymphocytic Leukemia

A doctoral study from the University of Liverpool shows that machine learning revolutionizes leukemia cell analysis, achieving 94% accuracy in gene expression prediction and opening new paths for precision medicine. The study combines FlowSOM clustering, XGBoost algorithm, etc., to solve the bottleneck of high-dimensional data processing in mass cytometry, facilitating subtype differentiation of chronic lymphocytic leukemia (CLL) and prediction of key gene expressions.

Mass cytometry can detect dozens of proteins at the single-cell level, but the data volume far exceeds the processing capacity of traditional methods. CLL has high clinical heterogeneity, and the mutation status of immunoglobulin genes is an important prognostic indicator, requiring fine phenotypic analysis and complex modeling.

The study developed multiple ML-driven methods covering batch effect correction, cell classification, etc. The ML method outperforms CytofRUV in batch effect correction, with p-values of 0.003 and 0.004 for anchor samples and validation samples respectively, effectively eliminating experimental bias.

Using the FlowSOM algorithm for clustering, clusters 10 and 1 were identified, with significant differences in their distribution between mutated (M-CLL) and unmutated (UM-CLL) samples. A model built based on 20 cluster features achieved 75% accuracy in subtype differentiation, proving that mutation status affects the expression of cell surface markers.

Using XGBoost to predict Ki67 and MYC mRNA expression levels, the accuracy reached 94% when integrating intracellular markers, and 80% even with only basic features. Feature importance analysis identified key molecules such as TCL1A and CD27, providing clues for mechanism understanding and therapeutic targets.

Compared with traditional statistical and manual analysis, ML methods can automatically learn high-dimensional patterns, integrate multi-source information, provide quantitative confidence and feature importance. Moreover, the standardized process is reproducible, and its generalizability can be extended to other single-cell data and disease studies.

It helps shift from descriptive diagnosis to predictive analysis. Clinicians can identify high-risk subtypes earlier, predict disease progression, evaluate treatment response, reduce reliance on traditional genetic testing, and lower costs. It will be applied in more disease fields in the future.