GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics

Kaito Shiku, Ahtisham Fazeel Abbasi, Ryoma Bise, Yuichiro Iwashita, Kazuya Nishimura, Andreas Dengel, Muhammad Nabeel AsimarXiv 2026 · arXiv

TL;DR

Predicting single-cell gene expression directly from standard H&E histology images is highly challenging because expression profiles vary wildly by cell type. This paper introduces GC-MoE, a mixture-of-experts framework that uses an image-based routing network to softly combine cell-type-specific expert models guided by genomic co-expression priors. By incorporating cell-type structures and local cell-to-cell interactions, GC-MoE sets a new state-of-the-art for single-cell spatial transcriptomics estimation without requiring cell-type labels at inference time.

What Problem Does This Paper Address?

Spatial transcriptomics (ST) maps gene expression within tissue architectures, providing crucial insights into disease progression, drug responses, and tissue function. However, physical single-cell ST technologies remain expensive and technically demanding. While predicting spot-level gene expression (which averages data across local regions containing multiple cells) from routine H&E images has advanced significantly, doing so at a true single-cell resolution remains a major hurdle.

At the single-cell level, morphologically similar or neighboring cells can exhibit vastly different gene expression profiles based on their specific cell type (e.g., neoplastic vs. inflammatory). Traditional single-predictor models fail because they try to map diverse cell-level targets to a single average profile, smoothing out distinct biological modes and failing to capture true cell-to-cell variability.

Figure 1: (a) Difficulty of Single-Cell ST Estimation: Despite subtle variations in single-cell mor- phology, gene expression exhibits highly diverse patterns. (b) Genomic knowledge: Single-cell gene expression is known to exhibit distinct patterns across different cell types. (c) Motivation of proposed method: Mixture-of-Experts Framework Using Cell-Type-Specific Expert Models.

Methods

The authors propose the Genomics-Guided Cell-Type-Specific Mixture of Experts (GC-MoE) framework. Instead of relying on a single shared regression model, GC-MoE dynamically routes cell features to specialized experts tailored to specific cell types.

Overview of the proposed GC-MoE architecture, highlighting the routing network, C2CA, and CAP modules.

The architecture consists of three key components:

  1. Dynamic Expert Assignment (Routing Network): Cell-centered H&E image features are extracted using a frozen pathology foundation model (CONCH). A routing network estimates cell-type probabilities from these features and softly blends the outputs of multiple cell-type-specific experts. During training, the router is supervised by automated cell-type annotations, but at inference, it operates entirely unsupervised from the image data alone.
  2. Cell-Type-Specific Co-Expression-Aware Predictor (CAP): To ensure each expert understands the unique gene-gene relationships of its designated cell type, CAP builds a cell-type-specific gene co-expression graph using external single-cell RNA-sequencing (scRNA-seq) data. A Graph Convolutional Network (GCN) processes this graph to generate gene-wise prediction parameters, regularized to ensure co-expressed genes share similar representations.
  3. Cell-to-Cell Interaction Attention (C2CA): A lightweight cross-attention module refines a target cell's features by incorporating contextual cues from neighboring cells, helping resolve local morphological ambiguities.

Datasets

The model was evaluated on three single-cell ST datasets measured via Xenium technology:

  • 10x Xenium: A breast cancer sample containing 5,508 cells.
  • COAD: A colorectal adenocarcinoma dataset from the HEST-1k benchmark, containing 830,644 cells across four slides.
  • IDC: An invasive ductal carcinoma dataset from HEST-1k, containing 1,735,885 cells across four slides.

External scRNA-seq data from the Gene Expression Omnibus (GEO) was leveraged to construct the cell-type-specific co-expression graphs for breast and colorectal cancer patients.

Key Insights and Ideas

  • Cell Type as a Natural Regularizer: Instead of treating single-cell ST prediction as a massive, homogeneous regression problem, leveraging cell types as discrete "modes" prevents the model from predicting intermediate, biologically unrealistic averages.
  • Genomics-Guided Parameter Sharing: By using prior scRNA-seq data to structure the prediction weights via a GCN, the model forces co-expressed genes to share similar prediction parameters. This injects established biological constraints directly into a computer vision pipeline.
  • Context Matters: Cells do not exist in a vacuum. Incorporating local neighborhood context via cross-attention (C2CA) helps correct for ambiguous single-cell visual features by looking at surrounding cell types.

Key Results

GC-MoE consistently outperformed both single-cell baselines and adapted spot-level models across all datasets. Performance was evaluated using the Pearson Correlation Coefficient (PCC) on the top 50 highly variable genes:

  • On 10x Xenium, GC-MoE achieved a PCC of 0.357, outperforming the strongest baseline (Single-Cell TRIPLEX* at 0.346).
  • On COAD, it achieved a PCC of 0.233 (vs. TRIPLEX* at 0.219).
  • On IDC, it reached a PCC of 0.191 (vs. TRIPLEX* at 0.184).

Qualitative t-SNE visualizations confirmed that GC-MoE successfully preserves distinct, tightly clustered cell-type-specific expression distributions compared to single-predictor baselines, which produce overly dispersed and blurred outputs.

Limitations and Future Work

The primary limitation acknowledged by the authors is dropout noise. Single-cell ST data inherently suffers from weaker signals and higher rates of missing data (zero counts) compared to spot-level data. While the CAP module partially mitigates this by leveraging gene co-expression networks, it remains a partial fix.

The authors suggest that future work should focus on developing more explicit dropout-handling mechanisms and exploring multi-modal training strategies that incorporate more stable spot-level or bulk RNA-seq measurements as auxiliary training signals.

Summary updated

Abstract

Histology-based single-cell spatial transcriptomics (ST) estimation aims to predict gene expression for individual cells from histopathological images and cell locations, reducing the need for costly single-cell ST measurements. Unlike existing histology-to-ST methods that mainly predict spot-level profiles for local regions containing multiple cells, this task requires modeling cell-to-cell expression variability, which is strongly structured by cell type. We propose Genomics-Guided Cell-Type-Specific Mixture-of-Experts (GC-MoE), which estimates cell-type probabilities with a routing network and softly combines cell-type-specific experts for gene expression prediction. To further encode cell-type-dependent gene programs, we introduce the Cell-Type-Specific Co-Expression-Aware Predictor (CAP), together with a lightweight Cell-to-Cell Interaction Attention (C2CA) module for neighboring-cell context. Experiments and ablations on public single-cell ST datasets show consistent improvements over existing single-cell and adapted spot-level baselines.
BibTeX
@misc{shiku2026gc,
  author = {Kaito Shiku and Ahtisham Fazeel Abbasi and Ryoma Bise and Yuichiro Iwashita and Kazuya Nishimura and Andreas Dengel and Muhammad Nabeel Asim},
  title = {{GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics}},
  year = {2026},
  eprint = {2606.02424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.02424},
}

Loading related content...

Figure