Reproduce “HistoPrism: Unlocking Functional Pathway Analysis from Pan-Cancer Histology via Gene Expression Prediction” on OpenResearch

About this paper

Susu Hu, Qinghe Zeng, Nithya Bhasker, Jakob Nikolas Kather, Stefanie Speidel

HistoPrism introduces a transformer-based model to predict spatial gene expression from H&E histology images across various cancer types, achieving superior biological coherence in gene pathway prediction and higher clustering utility compared to previous methods. The model also demonstrates significant improvements in computational and data efficiency.

Key insights

• Direct-mapping architectures can efficiently predict pan-cancer spatial gene expression from H&E images, delivering superior biological coherence, especially for low-variance functional pathways.
• The Gene Pathway Coherence (GPC) benchmark provides a more clinically relevant and biologically interpretable measure of model performance than traditional highly variable gene (HVG) correlations.
• Integrating global cancer type information via cross-attention effectively modulates local patch representations, improving overall predictive performance and generalization across diverse cancer types.

Summary by alphaXiv.