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

Abstract

We propose a method for estimating probability distributions over single cells, which we apply to fine-scale cellular deconvolution, which quantifies the composition of external bulk RNAseq samples at high resolution (i.e. at the single-cell or neighborhood level). Our method is based on a computationally-efficient convex optimization problem, and is also an application of the Generalized Cross Entropy method for density estimation. Our method has a much higher resolution than traditional approaches that require computing gene expression profiles at the cell-type level, and also compares favorably to recent high-resolution cellular deconvolution methods, with orders-of-magnitude speedup in computational efficiency. We implement this method in a Python package quipcell, available athttps://github.com/genentech/quipcell.