Gradient boosting regression and convolution improve deconvolution of bulk transcriptomes

Abstract

Bulk cell type deconvolution aims to estimate cell type composition from bulk transcriptomic data. So-called pseudobulk simulation, where single-cell RNA-seq data is aggregated to bulk-like expression profiles, represents a central concept in training and testing of deconvolution tools. However, deconvolution methods often lack interpretability and struggle to generalize from simulated to real data. We present GrooD (GradientBoostedDeconvolution), a second-generation deconvolution tool that uses gradient boosted trees trained on pseudobulks from scRNA-seq references. GrooDs pseudobulk simulations account for donor and condition variability to better model the complexity of real-world transcriptomes. We show that GrooD achieves state-of-the-art and superior deconvolution performance on human blood biospecimen. Furthermore, Grood provides visualizations of feature loadings and deconvolution results, that allow mechanistic insights for biological interpretation. We further integrate a convolution framework to assess the transcriptomic similarity between bulk and pseudobulk data, showing that higher similarity can indicate better deconvolution performance. GrooD deconvolution of blood transcriptomes from a large sepsis patient cohort identifies meaningful shifts in immune cell type composition that are associated with disease severity. By combining interpretability, robustness, and heterogeneous pseudobulk simulation, GrooD represents a powerful, user-friendly second-generation tool for cell type deconvolution.