Skip to content
Book Open access

STMoE: Multi-Scale Mixture-of-Experts for Single-Cell Gene Expression Prediction from Histology

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 75 references

Abstract

Predicting single-cell molecular state from routine histology could make spatial profiling more scalable, but existing histology-to- expression methods are mostly limited to spot-level targets, re- stricted gene panels, or organ-specific settings. This leaves three gaps for practical single-cell deployment: a lack of large aligned- cell resources, limited support for direct multi-organ prediction, and incomplete full-gene expression recovery from whole-slide images. We address these gaps by curating CellST-24M, a 23.94M- cell histology–transcriptomics resource spanning 66 studies, 25 organ categories, Visium HD, and Xenium, and by introducing STMoE, a WSI-only organ-routed prediction framework for full- gene single-cell expression inference. STMoE detects cells from H&E slides, encodes cell-centered and tissue-context crops with a frozen UNI2-h pathology foundation model, aligns image and expression embeddings with soft contrastive learning, and predicts full-gene profiles through organ-routed retrieval from training-only reference banks. On matched human lung benchmarks, STMoE achieves state-of-the-art cell-level performance, outperforming cell- adapted BLEEP, sCellST and GHIST under both in-domain and cross-patient settings. Across 22 organs, STMoE supports direct multi-organ prediction and achieves gene-wise Pearson up to 0.6451 and cell-wise Pearson up to 0.8379, while retaining approximately 60% predictive capability under cross-patient transfer. In down- stream marker-derived analysis, predicted profiles recover coarse cell-state structure with 0.5935 accuracy and 0.5922 balanced accu- racy. Together, CellST-24M and STMoE establish a benchmark and modeling framework for WSI-only, multi-organ, full-gene single- cell histology-to-expression prediction.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.