Clustering-independent estimation of cell abundances in bulk tissues using single-cell RNA-seq data

Abstract

Single-cell RNA-sequencing has transformed the study of biological tissues by enabling transcriptomic characterizations of their constituent cell states. Computational methods for gene expression deconvolution use this information to infer the cell composition of related tissues profiled at the bulk level. However, current deconvolution methods are restricted to discrete cell types and have limited power to make inferences about continuous cellular processes like cell differentiation or immune cell activation. We present ConDecon, a clustering-independent method for inferring the likelihood for each cell in a single-cell dataset to be present in a bulk tissue. ConDecon represents an improvement in functionality and accuracy with respect to current deconvolution methods. Using ConDecon, we discover the implication of neurodegenerative microglial inflammatory pathways in the mesenchymal transformation of ependymoma, recapitulate spatial patterns of cell differentiation during zebrafish embryogenesis, and make temporal inferences from bulk ATAC-seq data. Overall, ConDecon significantly enhances our understanding of dynamic cellular processes within bulk tissue samples.

Fine-scale cellular deconvolution via generalized maximum entropy on canonical correlation features

“…Memory usage quipcell 7 seconds < 16 GB CPM (Frishberg et al, 2019) 55406 seconds (15.4 hours) < 16 GB ConDecon (Aubin et al, 2023) 6654 seconds (1.85 hours) 120-220 GB Table 1: Comparison of runtime and memory usage to perform deconvolution on the 5 validation samples. Memory usage is approximate: quipcell and CPM were run in jobs that were allocated 16 GB memory, but in fact use much less memory.…”

“…Note that setting celltypes with negative probability to 0 may be the culprit for the bottom horizontal stripe in the CPM panels of Figure 2. Note also that CPM does make use of a lower-dimensional embedding of the reference data, but only to subsample the reference single-cells according to a grid, and this embedding is Figure 2: Comparison of our density estimation method with CPM (Frishberg et al, 2019) and ConDecon (Aubin et al, 2023). For each sample in the held-out validation study (Travaglini et al, 2020), we compare the estimated fraction of UMIs (top) or cells (bottom) coming from each celltype across several annotation levels (higher level means finer resolution).…”

“…Comparison of our density estimation method with CPM (Frishberg et al, 2019) and ConDecon (Aubin et al, 2023). For each sample in the held-out validation study (Travaglini et al, 2020), we compare the estimated fraction of UMIs (top) or cells (bottom) coming from each celltype across several annotation levels (higher level means finer resolution).…”

“…We propose a novel method for bulk deconvolution, quipcell, that is a convex optimization problem and a Generalized Cross Entropy method (Botev and Kroese, 2011). Quipcell represents each sample as a proba-bility distribution over some reference single-cell dataset, similar to the MELD differential abundance method (Burkhardt et al, 2021), or the ConDecon bulk deconvolution method (Aubin et al, 2023). However, quipcell fits this distribution using an alternative approach that only relies on solving an efficient convex optimization problem, instead of constructing nearest neighbor graphs or fitting predictive machine learning models.…”

“…However, CPM was designed to distinguish continuous cell states within cell types, and may not work well on datasets containing very diverse cell types. More recent high-resolution deconvolution methods include ConDecon (Aubin et al, 2023), which learns a map from rank-correlation scores to a probability distribution over single cells; and MeDuSa (Song et al, 2023), which deconvolves bulk samples along a linear continuous pseudotime trajectory using a mixed effects model. In addition, spatial deconvolution methods, such as Tangram (Biancalani et al, 2021), bear some similarities to high-resolution bulk deconvolution, but operate in a different regime with distinct challenges (many voxels with few cells per voxel) as well as advantages (additional spatial information, well-matched single-cell and spatial samples).…”