Slide-Omics: An Interpretable Bi-Directional Attention Framework for Integrating Multi-Omics and Pathology in Cancer Survival Analysis

Integrating histopathology with molecular profiles can enhance accuracy in cancer survival prediction, yet most existing multimodal methods treat molecular inputs as unstructured predictors rather than biologically organized entities. As a result, the learned cross-modal associations are difficult to interpret at the pathway level, and the biological significance of image–omics relationships remains obscure. We present Slide-Omics, a multimodal deep learning framework that jointly models whole-slide morphology and pathway-structured multi-omics data to yield interpretable survival estimates. Central to Slide-Omics is a bidirectional cross-attention mechanism: slide-to-omics attention identifies the biological pathways most informative given the observed tissue architecture, whereas omics-to-slide attention identifies the morphological patterns most informative given pathway activity. These complementary views are merged into a unified importance score for each KEGG pathway, producing an interpretable, pathway-level map linking phenotype to genotype. Our work makes two main contributions. First, we introduce bidirectional cross-attention as a principled strategy for modeling interactions between tissue morphology and molecular pathways, enabling cross-modal reasoning absent from prior multimodal survival models. Second, we biologically validate the derived pathway importance scores, demonstrating concordance with established disease mechanisms across multiple cancer types. Whole-slide representations are obtained either from a CNN encoder via HipoMap or from a Vision Transformer foundation model via UNI2, while multi-omics data—gene expression, DNA methylation, and copy number alterations—are structured into KEGG pathway matrices. Across four TCGA cohorts, Slide-Omics achieves competitive survival prediction performance while offering pathway-level interpretability that directly supports mechanistic insight and clinical decision-making. The source code and data are available at https://github.com/aimlx-labcpp/Slide-Omics.