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

Researchers introduce GC-MoE, a machine learning framework that predicts individual cell gene expression from histopathology images and spatial data, addressing limitations of existing methods that only work at the spot level. The approach combines cell-type-specific expert models with genomic guidance to capture cellular expression variability more accurately than current baselines.

GC-MoE represents a significant advancement in computational biology by tackling single-cell prediction from histological images—a problem that bridges microscopy and transcriptomics without requiring expensive spatial sequencing. Traditional spot-level methods aggregate multiple cells together, losing critical cell-to-cell variation information that is fundamentally determined by cell type identity. This research recognizes that gene expression patterns follow cell-type-dependent programs and uses a mixture-of-experts architecture where a routing network probabilistically assigns cells to specialized prediction models.

The innovation builds on established machine learning patterns—mixture-of-experts frameworks have proven effective in large language models and computer vision—now adapted for spatial biology. By incorporating a Cell-Type-Specific Co-Expression-Aware Predictor (CAP) module, the method encodes known gene co-regulation patterns, grounding predictions in biological reality rather than pure statistical learning. The addition of cell-to-cell interaction attention captures how neighboring cell types influence individual gene expression through paracrine signaling and microenvironment effects.

For the biotech and computational biology sectors, this advancement could substantially reduce costs associated with single-cell spatial transcriptomics workflows, which remain expensive and labor-intensive. Researchers could potentially generate cell-level expression predictions from routine histopathology slides already collected during clinical diagnosis. This democratizes access to transcriptomic information across pathology labs worldwide. The consistent improvements demonstrated on public datasets suggest the method generalizes well, positioning it for rapid adoption in research pipelines and eventually clinical applications focused on understanding tumor microenvironments and tissue organization.

• →GC-MoE predicts individual cell gene expression from histology images, improving upon existing spot-level methods that cannot resolve single-cell variation.
• →The mixture-of-experts framework with cell-type-specific models captures expression patterns structured by cell identity and genomic co-regulation rules.
• →Integration of cell-to-cell interaction attention modules models paracrine signaling effects from neighboring cells on gene expression.
• →Computational approach could substantially reduce costs of single-cell spatial transcriptomics by leveraging existing histopathology slide images.
• →Consistent experimental improvements on public datasets demonstrate method generalization across diverse tissue types and cell populations.