Towards a Cytometry Foundation Model: Interpretable Sample-level Predictive Modelling via Pretrained Transformers.

Area of Science:

• Computational Biology
• Immunology
• Data Science

Background:

• Foundation models have revolutionized scientific data modeling, but flow cytometry analysis remains limited by marker variability and data homogeneity.
• Automated analysis in flow cytometry is often bottlenecked by the need for fixed marker panels and homogeneous datasets, hindering scalability and generalizability.

Purpose of the Study:

• To introduce the Generalised Pretrained Cytometry Transformer (GPCT), an interpretable framework for flow cytometry data analysis.
• To develop a cytometry-specific pretraining regime enabling learning from heterogeneous marker panels for sample-level predictive modeling.

Main Methods:

• Developed the Generalised Pretrained Cytometry Transformer (GPCT), a novel framework for flow cytometry.
• Implemented a cytometry-specific pretraining strategy to learn transferable cellular representations.
• Evaluated GPCT on diverse datasets for classification accuracy and interpretability.

Main Results:

• GPCT achieves high classification accuracy across diverse flow cytometry datasets.
• Pretraining significantly enhances performance on data-scarce downstream tasks.
• The framework maintains interpretability, identifying influential cell subsets for biological validation.

Conclusions:

• GPCT represents a significant advancement towards a foundation model for flow cytometry.
• The model's interpretability facilitates biological validation and refinement of traditional gating strategies.
• GPCT offers a scalable and generalizable approach to analyzing high-dimensional cellular data.