Unsupervised clustering reveals cell types in 40 specimen aGVHD PBMC cytometry dataset

Key Findings

Our study highlights how unsupervised clustering of high-dimensional cytometry data can reveal clinically relevant immune subsets that are overlooked by traditional gating. Applying this approach to an 11-million-cell mass cytometry dataset, we identified a KLRG1⁺ TBET⁺ CD4⁺ T cell population that was missed by conventional gating strategies. When we compared functional marker expression between responders and non-responders to an aGVHD therapy, this subset showed significantly lower levels of CD25 and PD-1 among responders.

Overview

Traditional gating strategies in cytometry often fail to identify rare or unconventional immune cell subsets, especially in high-dimensional datasets. In this poster, we present a novel unsupervised clustering pipeline that we applied to an 11 million cell dataset derived from peripheral blood mononuclear cells (PBMCs).

In the study, we processed 40 PBMC specimens from 20 acute Graft-versus-Host Disease (aGVHD) patients—treated with a urinary-derived human chorionic gonadotropin/epidermal growth factor and standard immunosuppressive therapy—using a 44-marker mass cytometry panel.

Methods

• Gating: We processed 40 PBMC specimens with a 44-marker mass cytometry panel and manually gated the data using the scheme below.

‍

‍

• Clustering: We used Fast-PhenoGraph (FastPG) to cluster cells by first identifying K-nearest neighbors for each cell. We then calculated a Jaccard index to assign a weight to each pair of neighboring cells, measuring how strongly they were associated. Lastly, the Louvain algorithm was used to optimize cluster assignments based on the strength of these connections
• Cluster Naming: Teiko's naming algorithm compares marker expression in the cell clusters to manually gated populations using a cosine similarity index. We assign biologically relevant names to the clusters and validate them through visual inspection. This process includes reviewing a cluster-versus-marker heatmap and a series of UMAP plots—one for each cluster and marker—where we highlight only the selected cluster or marker.

‍

Finding the Novel KLRG+ TBET+ CD4+ T Cell Subset

We uncovered a novel CD4+ T cell subset (Cluster 26) with elevated KLRG1 and TBET expression—markers linked to T cell exhaustion and effector differentiation—that conventional gating missed.

After clustering and naming each of the cell populations, we visually inspect clusters to confirm that marker expression matches the assigned identity. We began by zooming in on Cluster 26 within the CD4+ T cell island.

‍

‍

Based on partial CD45RA expression and lack of CD27, our algorithm initially labeled Cluster 26 as a CD4+ TEMRA cell. We generated a cluster-versus-marker heat map to evaluate marker expression within the CD4⁺ T cell island and verify the cluster's assigned name.

‍

‍

To verify the assignment, we first examined CD3 and CD4 expression to confirm that Cluster 26 belonged to the CD4⁺ T cell compartment. We then assessed CD27 and CD45RA, two markers commonly used to define T cell subsets. While Cluster 26 showed low CD27 expression—consistent with central memory cells—its CD45RA expression was variable.

To explain this discrepancy, we evaluated additional markers and found that Cluster 26 expressed higher levels of TBET and KLRG1 compared to other CD4⁺ populations. This expression pattern points to a novel T cell subset that traditional gating would likely overlook.

‍

‍

Evaluating Siginificant Changes in Marker Expression Among Responders

After defining the KLRG1⁺ TBET⁺ CD4⁺ T cell subset, we compared its functional marker expression between responders and non-responders to the uhCG/EGF combination therapy.

We found that responders had decreased expression of CD25, compared to non-responders, within the cluster at the Baseline timepoint. These functional differences suggest a potential link between this novel population and therapeutic outcomes.

‍

‍