Global tree encoding of atlas-scale single-cell genomics

Abstract

The rapid expansion of single-cell genomic datasets has led to the compilation of biological resources comprising hundreds of millions of cells across tissues, developmental stages, and disease states. This has underscored the need for scalable and interpretable data representations that preserve the complex relationships and multi-scale organization of cellular states, while remaining computationally tractable at atlas scale. Existing approaches based on discrete abstractions have enabled cell annotation, clustering, and trajectory inference, but are often optimized for local inference tasks and may obscure continuous cellular relationships and multi-resolution structure within complex transcriptional and other genomic landscapes. Moreover, increasing dataset sizes often require information-reduction strategies such as random downsampling, limiting the resolution of rare cell populations and heterogeneous cellular states. Here, we present MILK, a scalable computational framework that organizes high-dimensional single-cell populations into unified tree representations. Across large-scale transcriptomic atlases, MILK enables representative subsampling with preserved information, supporting the tractable application of existing algorithms for tasks including deep generative model training and foundation model benchmarking. Additionally, MILK enables holistic, multi-resolution analyses that capture global developmental trajectories, characterize disease-associated cellular perturbations across tissues, and facilitate comparison of transcriptional programs across species within a coherent hierarchical framework. Together, these results establish the hierarchical organization of biological data as a scalable and unifying representation of cellular identity, enabling integrative analysis of single-cell genomic data across diverse contexts.