# HIDF: Integrating Tree‐Structured scRNA‐seq Heterogeneity for Hierarchical Deconvolution of Spatial Transcriptomics

## TL;DR

HIDF improves spatial transcriptomics by modeling hierarchical cell types, revealing subtype-level spatial patterns missed by existing methods.

## Contribution

HIDF introduces a hierarchical iterative framework with dual regularization to model cell type hierarchies in spatial transcriptomics.

## Key findings

- HIDF outperforms existing methods on simulated and real spatial transcriptomics datasets.
- HIDF reveals subtype-level spatial heterogeneity undetectable by conventional approaches.
- The method provides accurate cell type distributions aligned with known tissue functions.

## Abstract

The limited spatial resolution of mainstream spatial transcriptomic technologies captures transcriptomic mixtures from multiple cells per spot, obscuring crucial single‐cell information. While numerous methods leverage single‐cell RNA sequencing references to infer cellular composition from ST data, they primarily rely on fixed cell type labels, overlooking the intrinsic hierarchical heterogeneity (subtypes within broad types) of cellular populations and its association with spatial organization. To address this limitation, HIDF, a Hierarchical Iterative Deconvolution Framework is proposed. HIDF progressively resolves cellular heterogeneity from coarse to fine granularity, it employs a hierarchical iterative optimization mechanism guided by the cluster‐tree to recover single‐cell spatial distributions. This process is further stabilized and enhanced by incorporating dual regularization constraints (spatial neighborhood and cross‐level regularization). Comprehensive benchmarking demonstrates that HIDF outperforms existing methods on simulated and real tissue datasets. In addition, HIDF not only reveals cell type distributions consistent with known tissue functions but also uncovers spatially heterogeneous patterns of cell subtypes undetectable by conventional methods.

The prevailing neglect of cellular hierarchies in current spatial transcriptomics deconvolution often obscures cellular heterogeneity and impedes the identification of fine‐grained subtypes. To address this issue, HIDF employs a cluster‐tree and dual regularization to systematically model cellular hierarchical structures. This approach significantly improves deconvolution accuracy, achieves precise subtype‐level resolution, and provides a foundation for a deeper understanding of tissue spatial organization.