Learning discriminative prototypes via Variational Information Bottleneck for multimodal cancer survival prediction

Survival prediction is critical in cancer research for forecasting patient outcomes and optimizing treatment strategies. While integrating histopathology whole-slide images (WSIs) with genomic profiles holds promise for improved predictive accuracy, two major challenges persist: (1) the inherent low signal-to-noise ratio in gigapixel WSIs, where task-irrelevant background patches often dominate standard Multi-Instance Learning (MIL) bags, and (2) the instability of unsupervised prototype initialization (e.g., K-means), which frequently converges to non-informative morphological patterns. To address these challenges, we propose DP-VIB (Discriminative Prototypes via Variational Information Bottleneck), a hierarchical framework that leverages the information bottleneck principle to learn survival-relevant prototypes in a coarse-to-fine manner. Unlike traditional approaches, DP-VIB first employs a survival-supervised Global VIB Module to learn a discriminative latent representation, filtering out nuisance variations through an information bottleneck constraint. Subsequently, Gaussian Mixture Model (GMM) priors derived from this purified latent representation serve as robust initializations for fine-grained patch-level prototypes, significantly enhancing convergence and semantic quality. Finally, we introduce a tri-branch Histology-Genomics(HG) encoder followed by a cross-modal fusion mechanism to deeply fuse these VIB-guided prototypes with genomic pathway embeddings, enabling analysis of model-derived prototype-pathway associations. Extensive validation on five TCGA cancer cohorts demonstrates that DP-VIB consistently outperforms strong baselines in Concordance Index (C-index), achieving superior robustness and accuracy in individualized survival prediction.