Defining Cell Types through Maximally Informative Biological Atlases

Abstract

Cell type atlases organize biological complexity by compressing high-dimensional molecular profiles into discrete cell types that can be mapped across tissues. Yet a principled theory of what constitutes a cell type remains lacking. Here, I introduce an information-theoretic framework in which molecularly defined cell type classifications are evaluated by the information content of the spatial maps they produce. The optimal taxonomy maximizes spatial information, constrained by molecular data, by balancing the entropy of the cell type code against the spatial entropy of the resulting map. The framework generalizes naturally to higher-order spatial partitions, including tissue regions defined by local cellular composition. Applied to a whole-brain mouse spatial transcriptomic dataset, this approach identifies optimal cell-type and region-level taxonomies that balance coding complexity with spatial informativeness. Together, these results establish a unified information-theoretic foundation for cell-type and region-level atlas construction grounded in tissue spatial architecture.