Self-supervised deep learning enables label-free, high-dimensional morphology profiling of immune cell types

---

Presentation Time: 11:30 AM - 12:45 PM
Poster Board Number: B881
Abstract ID: 4681

Presenting Author:
Michelle Balakrishnan , Sr. Product Marketing Manager at Deepcell, Inc.

Abstract:

Immunophenotyping & detecting cell subsets are key parts of immunology research. To determine whether morphology as a quantitative, high-dimensional analyte (morpholome) can phenotype and distinguish among cell types found in PBMCs more comprehensively & without labels we used the REM-I platform to assess patient-derived PBMC samples & purified immune cell subsets.

The REM-I platform combines high-speed imaging in flow on the REM-I instrument, and deep learning & computer vision in the Human Foundation Model (HFM). The HFM produces 115 embeddings that are reproducible, quantitative descriptions representing the morpholome of each cell in a sample that can be analyzed through interactive UMAPs, population analytics, & image visualizations. 

Morpholomic analysis of patient-derived PBMC samples and purified immune cell subsets including CD4+ T cells, CD8+ T cells, CD14+ monocytes, CD19+ B cells, CD38+ plasma cells, CD56+ NK cells, and macrophages, showed profiled cell subsets in distinct locations on the associated UMAP, indicating cell types are separable by morpholomics. Leiden clustering segregated cell groups by morphological features, with some cell subsets sharing common leiden clusters. Purified cell types derived from different donors consistently localized to the same high-density UMAP regions, indicating that diverse morphotypes are reproducible across different donors, & suggesting a link between morphological profile and cell function.