Unsupervised machine learning reveals risk stratifying glioblastoma tumor cells

Abstract

<jats:p>A goal of cancer research is to reveal cell subsets linked to continuous clinical outcomes to generate new therapeutic and biomarker hypotheses. We introduce a machine learning algorithm, Risk Assessment Population IDentification (RAPID), that is unsupervised and automated, identifies phenotypically distinct cell populations, and determines whether these populations stratify patient survival. With a pilot mass cytometry dataset of 2 million cells from 28 glioblastomas, RAPID identified tumor cells whose abundance independently and continuously stratified patient survival. Statistical validation within the workflow included repeated runs of stochastic steps and cell subsampling. Biological validation used an orthogonal platform, immunohistochemistry, and a larger cohort of 73 glioblastoma patients to confirm the findings from the pilot cohort. RAPID was also validated to find known risk stratifying cells and features using published data from blood cancer. Thus, RAPID provides an automated, unsupervised approach for finding statistically and biologically significant cells using cytometry data from patient samples.</jats:p>

Mapping circuit dynamics during function and dysfunction

“…Here we used the t-SNE algorithm as a core method to reduce the dimensionality of the dataset and to visualize our data. t-SNE has been widely used in the unsupervised analysis of many types of biological data (Berman et al, 2014; Kollmorgen et al, 2020; Chen et al, 2020; Macosko et al, 2015; Kobak and Berens, 2019; Leelatian et al, 2020), including neural recordings (Dimitriadis et al, 2018). t-SNE minimizes the Kullback-Leibler divergence between a Gaussian distribution modeling pairwise distances between data points and a Student t-distribution modeling distances between the same points in a low (typically two) dimensional embedding (Van der Maaten and Hinton, 2008; Linderman and Steinerberger, 2019).…”

Unsupervised machine learning reveals key immune cell subsets in COVID-19, rhinovirus infection, and cancer therapy

“…Marker Enrichment Modeling from the MEM package (https://github.com/cytolab/mem) was used to characterize feature enrichment in KNN region around each cell. MEM normally requires a comparison of a population against a reference control, such as a common reference sample (Diggins et al, 2017), all other cells (Diggins et al, 2018;Leelatian et al, 2020), or induced pluripotent stem cells (Greenplate et al, 2019). Here, a statistical reference point intended as a statistical null hypothesis was used as the MEM reference.…”

“…Analysis algorithms typically rely on aggregate statistics for groups of cells, but the process of grouping the cells works best with larger, established populations ( Diggins et al, 2015 ; Irish et al, 2006 ; Saeys et al, 2016 ) or may include pre-filtering of cells by human experts ( Greenplate et al, 2016a ; Greenplate et al, 2019 ). Cytometry tools like SPADE ( Bendall et al, 2011 ; Qiu et al, 2011 ), FlowSOM ( Van Gassen et al, 2015 ), Phenograph ( Levine et al, 2015 ), Citrus ( Bruggner et al, 2014 ), and RAPID ( Leelatian et al, 2020 ) generally work best to characterize cell subsets representing >1% of the sample and are less capable of capturing extremely rare cells or subsets distinguished by only a fraction of measured features. Tools like t-SNE ( Amir el et al, 2013 ; Krijthe et al, 2015 ), opt-SNE ( Belkina et al, 2019 ), and UMAP ( Becht et al, 2018 ; McInnes et al, 2018 ) embed cells or learn a manifold and represent these transformations as algorithmically-generated axes.…”

“…Tools like t-SNE ( Amir el et al, 2013 ; Krijthe et al, 2015 ), opt-SNE ( Belkina et al, 2019 ), and UMAP ( Becht et al, 2018 ; McInnes et al, 2018 ) embed cells or learn a manifold and represent these transformations as algorithmically-generated axes. In addition to assisting with data visualization, these tools frequently reveal unexpected cells and facilitate their identification through manual or automated clustering ( Amir el et al, 2013 ; Becher et al, 2014 ; Diggins et al, 2015 ; Diggins et al, 2017 ; Gandelman et al, 2019 ; Leelatian et al, 2020 ). Sconify ( Burns et al, 2018 ) is one such tool that applies k -nearest neighbors (KNN) to calculate aggregate statistics for the immediate phenotypic neighborhood around a given cell on a t-SNE plot that combines data from multiple cytometry samples.…”

“…For a biologist, these tools provide a way to organize cells according to phenotypic relationships that span multiple measured features, such as the proteins quantified on each of millions of cells in the datasets here. In addition to assisting with data visualization, these tools frequently reveal unexpected cells and facilitate their identification through manual or automated clustering 6, 15, 16, 21–23 . Sconify 24 is one such tool that applies k -nearest neighbors (KNN) to calculate aggregate statistics for the immediate phenotypic neighborhood around a given cell on a t-SNE plot representing data from multiple cytometry samples.…”