Dissecting heterogeneous cell populations across drug and disease conditions with PopAlign

Abstract

Single-cell measurement techniques can now probe gene expression in heterogeneous cell populations from the human body across a range of environmental and physiological conditions. However, new mathematical and computational methods are required to represent and analyze gene-expression changes that occur in complex mixtures of single cells as they respond to signals, drugs, or disease states. Here, we introduce a mathematical modeling platform, PopAlign, that automatically identifies subpopulations of cells within a heterogeneous mixture and tracks gene-expression and cell-abundance changes across subpopulations by constructing and comparing probabilistic models. Probabilistic models provide a low-error, compressed representation of single-cell data that enables efficient large-scale computations. We apply PopAlign to analyze the impact of 40 different immunomodulatory compounds on a heterogeneous population of donor-derived human immune cells as well as patient-specific disease signatures in multiple myeloma. PopAlign scales to comparisons involving tens to hundreds of samples, enabling large-scale studies of natural and engineered cell populations as they respond to drugs, signals, or physiological change.

Machine learning for perturbational single-cell omics

“…We provide a review of existing single-cell omics perturbation models separated into categories based on commonly established ways to categorize ML models (Table 1). All methods except CellOracle (Kamimoto et al, 2020) can be used for perturbations as defined by a before-and-after effect, which could include healthy versus diseased phenotypes (Buschur et al, 2020), or cross-species translation (Lotfollahi et al, 2019;Chen et al, 2020aChen et al, , 2020b. CellOracle is specific to genetic/single-target perturbations as it infers effect through propagating signal through a gene regulation network (GRN).…”

“…While also applicable to bulk transcriptomics (Umarov and Arner, 2020;Rampá sek et al, 2019), distribution modeling gained popularity in the single-cell field as a way to describe population shifts and is especially tractable given the number of cells. PopAlign (Chen et al, 2020a(Chen et al, , 2020b fits a Gaussian and matches perturbed and unperturbed cell populations after factor decomposition into a latent space (with orthogonal non-negative matrix factorization, such that PopAlign is also in part a factor decomposition model). PhEMD The ''data'' column describes the data modality the model is based on as described in the original paper.…”

Active feature selection discovers minimal gene sets for classifying cell types and disease states with single-cell mRNA-seq data

“…We test ASFS method on three biology datasets: a dataset of peripheral blood mononuclear cells (PBMCs) [13], the Tabula Muris dataset [14], and MM datasets [15]. We test with both mincell strategy and min-complexity strategy introduced in Cell Selection part.…”

“…Having demonstrated the method on identifying cell-type specific markers at both small-and largescale, we next turned to applying ASFS to discovering disease-specific markers. We used singlecell data from peripheral blood immune cells collected from both healthy donors and patients who have been diagnosed with multiple myeloma [15]. Multiple myeloma (MM) is an incurable cancer of plasma cells, known as myeloma cells, that overproliferate in the bone marrow.…”

“…We apply our active feature selection method to three test cases: a dataset of peripheral blood mononuclear cells (PBMCs) [13], the Tabula Muris mouse tissue survey [14], and a data set of multiple myeloma patient PBMCs [15]. We also systematically compare the performance of the method to six existing conventional feature selection methods, showing that our method outperforms other methods in terms of classification accuracy.…”