MuSARCyto: Multi-Head Self-Attention-Based Representation Learning for Unsupervised Clustering of Cytometry Data.

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Cytometry is crucial for disease diagnosis and monitoring, but manual cell clustering (gating) is subjective and time-consuming.
• Existing automated solutions often underperform compared to manual gating, limiting their clinical adoption.

Purpose of the Study:

• To develop an improved unsupervised deep learning (DL) architecture for cytometry data clustering.
• To enhance the performance of automated cytometry data representation and cluster assignment.

Main Methods:

• Proposed MuSARCyto, a multi-head self-attention-based representation learning network (RN) for unsupervised cytometry clustering.
• Utilized a fully-connected representation network backbone for efficient data representation.
• Introduced the adjudicator score (Ad_n) as an ensemble metric to evaluate clustering performance.

Main Results:

• MuSARCyto demonstrated superior performance compared to state-of-the-art methods across six cytometry datasets.
• The proposed DL architectures are computationally efficient and deployable in clinical settings.
• The study validated the effectiveness of DL for identifying meaningful cell clusters in immunology.

Conclusions:

• MuSARCyto offers a robust and efficient solution for automated cytometry data clustering.
• Deep learning methods show significant potential to overcome the limitations of manual gating in clinical and research settings.
• This work paves the way for more widespread and accurate application of cytometry in disease analysis.