ProjectSVR: mapping single-cell RNA-seq data to reference atlases by supported vector regression.

Abstract

Mapping the query cells onto a well-constructed reference atlas, known as reference mapping, enables robust, reproducible interpretation of new single-cell RNA-seq data in the context of curated and annotated cell subtypes and states. However, existing methods often rely on complex integration frameworks or require re-access to raw data, limiting their applicability and reproducibility. To address this, we introduce ProjectSVR, a machine learning-based framework that formulates reference mapping as a multi-target regression task. By leveraging ensemble support vector regression (SVR) to learn the relationship between gene set activity scores and low-dimensional reference embeddings, ProjectSVR enables platform-agnostic and integration-independent mapping. Benchmarking across diverse biological contexts-including immune responses, developmental trajectories, and disease states-demonstrates that ProjectSVR achieves comparable accuracy and robustness to state-of-the-art methods, with reduced dependence on data-specific preprocessing. Our findings demonstrate that ProjectSVR is a valuable tool for reference mapping, considerably simplifying the analysis of scRNA-seq data when well-constructed reference atlases are available.