Title: GC-MoE: A Genomics-Guided Mixture of Experts Approach for Cell-Type-Specific Modeling in Histology-Based Single-Cell Spatial Transcriptomics

Abstract

The objective of histology-based single-cell spatial transcriptomics (ST) estimation is to infer gene expression profiles for individual cells using histopathological images alongside their spatial coordinates, thereby mitigating the high costs associated with single-cell ST measurements. While current histology-to-ST techniques primarily focus on predicting spot-level profiles for regions comprising multiple cells, this specific task necessitates modeling the variability in expression between individual cells—a variance that is heavily influenced by cell type. To address this, we introduce the Genomics-Guided Cell-Type-Specific Mixture-of-Experts (GC-MoE). This framework estimates the probability of cell types via a routing network and employs a soft combination of cell-type-specific experts to predict gene expression. Additionally, to better capture gene programs that depend on cell type, we propose the Cell-Type-Specific Co-Expression-Aware Predictor (CAP). This is complemented by a lightweight Cell-to-Cell Interaction Attention (C2CA) module, which incorporates context from neighboring cells. Evaluations and ablation studies conducted on public single-cell ST datasets demonstrate that our approach consistently outperforms both existing single-cell methods and adapted spot-level baselines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC