Boosting Single-Cell Methods Significantly Enhances Cell-type Classification

Single-cell sequencing drives biological discovery by revealing key cellular processes, cell types, and cell states. Recent technological advances have accelerated atlas-scale data generation across laboratories. With the increasing amount of single-cell RNA-seq (scRNA-seq) data, new challenges emerge, including effective batch correction, cell-type annotation, rare cell-type and cell-state detection, as well as biological pathway discovery. Inspired by methods for boosting weak learners such as AdaBoost, here we present a novel paradigm, scBoost, that combines single-cell analysis methods using stacked ensembles to boost task performance beyond that of individual methods. We show that scBoost pretrained on Tabula Sapiens 2.0 dataset performs significantly better than existing methods in cell-type classification on the real-world Human Fetal Heart Map and Plaque Atlas. We also demonstrate that scBoost substantially improves the performance of cell-type classification and ranks at the top of the leaderboard on the OpenProblems benchmark, a standard-in-the-field. scBoost is modular and is adaptive to new single-cell methods, which will enable it to continue to serve as a resource to the community in the future.