Loss-Function Learning for Digital Tissue Deconvolution

Abstract

The gene expression profile of a tissue averages the expression profiles of all cells in this tissue. Digital tissue deconvolution addresses the following inverse problem: given the expression profile y of a tissue, what is the cellular composition c of that tissue? If X is a matrix whose columns are reference profiles of individual cell types, the composition c can be computed by minimizing L(y-Xc) for a given loss function L. Current methods use predefined all-purpose loss functions. They successfully quantify the dominating cells of a tissue, while often falling short in detecting small cell populations. In this study we use training data to learn the loss function L along with the composition c. This allows us to adapt to applicationspecific requirements such as focusing on small cell populations or distinguishing phenotypically similar cell populations. Our method quantifies large cell fractions as accurately as existing methods and significantly improves the detection of small cell populations and the distinction of similar cell types.

Adaptive Digital Tissue Deconvolution

“…It was repeatedly shown that gene selection strongly affects the performance of bulk deconvolution methods [15, 10], and the importance of a stringent marker gene selection was thoroughly demon-strated [13]. To identify and weigh genes according to their importance for tissue dissection, we used artificial mixtures of single-cell RNA sequencing profiles from healthy breast tissue (see Methods).…”

“…(2) can be solved for each column (bulk) in Y , individually. In [10, 16], Eq. (2) was augmented by gene weights g = ( g 1 , g 2 , …, g p ) T , yielding where the weights g i can be chosen to improve estimates of C .…”

“…(2) was augmented by gene weights g = ( g 1 , g 2 , …, g p ) T , yielding where the weights g i can be chosen to improve estimates of C . They can be either specified by prior knowledge or can be learned from cellular mixtures with known cellular contributions, such as annotated bulk profiles or artificial mixtures generated from single-cell RNA sequencing data [10, 16]. To optimize the gene weights, we defined the outer loss function where C j · is the j th row (cell type) of C and Ĉ j · its corresponding estimate.…”

“…Single-cell RNA sequencing data can substantially improve estimates of cellular proportions via improved reference matrices as well as collections thereof, and via optimized gene selection or weighting, as demonstrated by work of us and others [8, 9, 10, 11, 12]. The latter is usually facilitated by generating artificial mixtures of known cellular composition which are used to assess the model performance by comparing the ground truth cellular composition with its estimate.…”

“…The latter is usually facilitated by generating artificial mixtures of known cellular composition which are used to assess the model performance by comparing the ground truth cellular composition with its estimate. Using such a strategy, for instance, MuSiC [9] and DTD [10] learn a gene weighting and Bisque [11] learns gene-specific bulk expression transformations.…”