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

Abstract

Recent developments in machine learning implemented dimensionality reduction and clustering tools to classify the cellular composition of patient-derived tissue in multi-dimensional, single cell studies. Current approaches, however, require prior knowledge of either categorical clinical outcomes or cell type identities. These algorithms are not well suited for application in tumor biology, where clinical outcomes can be continuous and censored and cell identities may be novel and plastic. Risk Assessment Population IDentification (RAPID) is an unsupervised, machine learning algorithm that identifies single cell phenotypes and assesses clinical risk stratification as a continuous variable. Single cell mass cytometry evaluated 34 different phospho-proteins, transcription factors, and cell identity proteins in tumor tissue resected from patients bearing IDH wild-type glioblastomas. RAPID identified and characterized multiple biologically distinct tumor cell subsets that independently and continuously stratified patient outcome. RAPID is broadly applicable for single cell studies where atypical cancer and immune cells may drive disease biology and treatment responses.

Two New Neutrophil Subsets Define a Discriminating Sepsis Signature

“…To identify immune subsets and visualize all cells in a 2D map where position represents local phenotypic similarity, we used two different dimensionality reduction tools depending on the strategy: the viSNE implementation of t-SNE (29) and the UMAP (30). Cells were also grouped in phenotypically homogenous clusters using either SPADE (31) or FlowSOM (32,33). To phenotypically characterize these clusters, Marker Enrichment Modeling (MEM) (34,35) was used.…”

Two new immature and dysfunctional neutrophil cell subsets define a predictive signature of sepsis useable in clinical practice

“…In the validation strategy, Uniform Manifold Approximation and Projection (UMAP) (25) was used to create a single, common map of neutrophils across all samples (18,800 cells x 67 samples) using the 7 markers previously used (Supplemental Table3). Once a common two-dimensional representation of all samples was established, a FlowSOM clustering optimization was applied to determine optimal cluster number based on relative homogeneity of marker expression in clusters (42). The self-organized map generated in an unsupervised way with the optimized FlowSOM algorithm produced clusters that were stable and contained phenotypically homogenous cells (26).…”

Multiplex bioimaging of single-cell spatial profiles for precision cancer diagnostics and therapeutics

“…Further, multiplexed single-cell proteomic profiles were measured in pediatric diseases and a machine-learning model (regularized elastic-net approach) was used to incorporate cellular developmental aberrations to predict the relapse of leukemias 113 . Similarly, a subset of glioblastoma cells were identified as the response predicter from single-cell signaling profiles without the need for prior information about the expected cell clusters on bioinformatics maps 114 . When combined with spatial multi-omics technologies, bioimaging features (cellular positions, interaction frequencies, tissue structural variations, and subcellular distributions) from single-cell analysis and molecular imaging will continue impacting predictive outcome studies in precision medicine.…”