Title: RankByGene: Achieving Cross-Modal Ranking Consistency in Gene-Guided Histopathology Representation Learning

Abstract: Spatial transcriptomics (ST) offers crucial spatial context by mapping gene expression within tissue structures, facilitating in-depth analysis of cellular heterogeneity and tissue organization. Nevertheless, integrating ST data with histology images is difficult due to inherent spatial distortions and variations specific to each modality. Current approaches predominantly depend on direct alignment, which frequently fails to capture intricate cross-modal relationships. To overcome these constraints, we introduce a novel framework that aligns gene and image features through a ranking-based alignment loss. This approach preserves relative similarity across modalities and facilitates robust multi-scale alignment. To further bolster the stability of this alignment, we utilize self-supervised knowledge distillation within a teacher-student network architecture. This strategy effectively counteracts disruptions arising from the high dimensionality, sparsity, and noise characteristic of gene expression data. Comprehensive experiments conducted on seven public datasets, covering gene expression prediction, slide-level classification, and survival analysis, highlight the effectiveness of our method. The results demonstrate superior alignment and predictive performance compared to existing techniques.

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