Computational Immune Monitoring Reveals Abnormal Double-Negative T Cells Present across Human Tumor Types

Abstract

Advances in single-cell biology have enabled measurements of >40 protein features on millions of immune cells within clinical samples. However, the data analysis steps following cell population identification are susceptible to bias, time-consuming, and challenging to compare across studies. Here, an ensemble of unsupervised tools was developed to evaluate four essential types of immune cell information, incorporate changes over time, and address diverse immune monitoring challenges. The four complementary properties characterized were: 1) systemic plasticity, 2) change in population abundance, 3) change in signature population features, and 4) novelty of cellular phenotype. Three systems immune monitoring studies were selected to challenge this ensemble approach. In serial biopsies of melanoma tumors undergoing targeted therapy, the ensemble approach revealed enrichment of double-negative (DN) T cells. Melanoma tumor resident DN T cells were abnormal and phenotypically distinct from those found in non-malignant lymphoid tissues, but similar to those found in glioblastoma and renal cell carcinoma. Overall, ensemble systems immune monitoring provided a robust, quantitative view of changes in both the system and cell subsets, allowed for transparent review by human experts, and revealed abnormal immune cells present across multiple human tumor types.

High risk glioblastoma cells revealed by machine learning and single cell signaling profiles

“…These patient-specific t-SNE maps were generated using 26 of the 34 measured markers 18 (Supplemental Table 2). Patient specific t-SNE maps revealed nonglioblastoma populations of immune (CD45 + ) and endothelial (CD45 -CD31 + ) cells, consistent with prior mass cytometry studies of gliomas 11,21,24 . Non-immune, non-endothelial cells were computationally isolated from each individual patient prior to subsequent downstream analysis of tumor-intrinsic phenotypic parameters (Figure 1, 2).…”

“…The second round of data analysis used an equal number of each patient's glioblastoma cells to create a single, common t-SNE map of glioblastoma cell phenotypes across all patients (N = 131,880 cells; 4,710 cells x 28 patients). Prior to creating this common map, mass cytometry standardization beads were used to remove batch effects and to set the variance stabilizing arcsinh scale transformation for each channel following field-standard protocols 11,38,41 . This common t-SNE map was generated using 24 of 34 measured markers (Supplementary Table 2) and was used for automated analysis of risk stratifying cell subsets.…”

“…Quantitative analysis of single cell cytometry data has recently moved from an era of human-driven identification of cell types using guide markers (expert gating) and embraced machine learning tools that can automatically reveal and characterize novel and abnormal cells [10][11][12][13] . In building an automated cytometry workflow, algorithm developers need to decide whether users will supervise the discovery of cell subsets using clinical knowledge.…”

“…Mass cytometry has recently been developed for human solid tumors 11,16,24,25 , including glioblastoma, the most common primary malignant brain tumor in adults 26 . The median survival of glioblastoma patients after diagnosis has remained approximately 12-15 months for over a decade 27,28 .…”

Deep immune profiling of COVID-19 patients reveals patient heterogeneity and distinct immunotypes with implications for therapeutic interventions

“…For each cell subset, a new self-organizing map (SOM) was generated using hierarchical consensus clustering on the tSNE axes. For each SOM, 100 clusters and 10 metaclusters were identified.To group individuals based on B cell landscape, pairwise Earth Mover's Distance (EMD) value was calculated on the B cell tSNE axes for all COVID-19 day 0 patients, healthy donors, and recovered donors using the emdist package in R as previously described(57). Resulting scores were hierarchically clustered using the hclust package in R.…”