ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction

Technical Summary of ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction

Problem Statement
Current multimodal survival prediction models for cancer often integrate whole slide images (WSIs) and bulk transcriptomic profiles but suffer from two primary limitations: a lack of inherent interpretability and the absence of biologically grounded semantic identity in their cross-modal fusion mechanisms. Existing methods frequently rely on post-hoc explainability techniques (e.g., SHAP, LIME) that treat models as black boxes, or they fuse modalities using representations that lack stable meaning across patients (e.g., attention over individual, non-consistent patches or independent pathway encoders that ignore gene-pathway topology). Clinicians require models where reasoning is transparent by design, with representations that align with established biological hierarchies to validate predictions against known disease mechanisms.

Methodology
ProtoPathway is an interpretable-by-design framework that unifies WSI and transcriptomics through encoders producing biologically grounded representations on both sides of the fusion. The architecture consists of three core components:

1. Learnable Morphological Prototypes (WSI Encoder): Instead of aggregating thousands of variable-length image patches via attention weights, ProtoPathway compresses the WSI into a fixed set of $K$ learnable prototype vectors. These prototypes are trained end-to-end with the survival objective. Patches are assigned to prototypes via soft assignment based on cosine similarity in a shared projected space. The resulting prototype tokens serve as the slide representation itself, capturing recurring morphological patterns relevant to prognosis. This allows for spatial interpretability, as patches can be mapped back to specific tissue locations based on their assigned prototype.

2. Topology-Aware Transcriptomic Encoding (Genomic Encoder): Rather than using independent Self-Normalizing Networks (SNNs) for each pathway, ProtoPathway employs a bipartite Graph Neural Network (GNN) over the Reactome pathway hierarchy. Genes and pathways form two node types connected by bidirectional membership edges. Information propagates through GraphSAGE layers with mean aggregation, allowing genes participating in multiple pathways to facilitate implicit cross-pathway communication. A final GATv2 attention layer computes gene importance weights within pathways. This structure ensures that pathway embeddings reflect both constituent genes and their broader biological context.

3. Semantically Grounded Asymmetric Cross-Modal Fusion: The fusion mechanism models the biological causal direction from molecular programs to tissue morphology. Prototype tokens (morphology) query pathway embeddings (genomics) via multi-head cross-attention. This produces a compact $K \times P$ attention matrix where prototypes act as queries and pathways as keys/values. Because both axes carry stable, task-learned identity, this attention matrix is not merely an internal computation but a direct interpretability output, linking specific tissue patterns to specific biological programs.

Key Contributions
The paper introduces three specific contributions:

• Learned Prototypes as Slide Representation: The model uses $K$ learnable prototype vectors trained end-to-end as the primary slide representation, rather than as auxiliary reference points. These tokens serve dual roles as morphological aggregators and as queries in cross-modal attention.
• Bidirectional Message Passing in Transcriptomics: A bipartite GNN encodes gene expression within the Reactome hierarchy, enabling parameter sharing and cross-pathway communication through shared genes, replacing the independent per-pathway encoders used in prior work.
• Asymmetric Fusion with Native Interpretability: The cross-attention mechanism operates over a prototype $\times$ pathway matrix, reflecting the biological direction from molecular programs to morphology. This design yields a traceable attribution chain from genes through pathways and prototypes to spatial tissue maps without requiring post-hoc methods.

Experimental Results
ProtoPathway was evaluated on five TCGA cancer cohorts (BRCA, BLCA, COADREAD, HNSC, STAD).

• Performance: The model achieved the highest overall Concordance Index (C-index) of 0.670, outperforming established multimodal baselines such as MCAT (0.662) and MMP (0.659). It ranked first on four of the five cohorts. Notably, on the HNSC cohort, it was the only multimodal method to clearly surpass the gene-only SNN baseline.
• Efficiency: The prototype-based compression significantly reduced computational costs. ProtoPathway required 480K parameters and 3.9G FLOPs, achieving a 28–50$\times$ training speedup compared to methods like MCAT and SurvPath, while using less VRAM (325 MB vs. 818 MB for SurvPath).
• Interpretability: Qualitative analysis on the BLCA cohort demonstrated that the model recovered known prognostic signatures. For instance, it identified that muscle invasion (a key staging factor) and stromal-immune composition were critical, shifting attention away from morphologically dominant tumor regions toward the microenvironment after incorporating pathway context. Spatial overlays successfully mapped specific pathways (e.g., cholesterol metabolism, cytokine signaling) and genes (e.g., IL7R, FGFR2b) to distinct tissue compartments.

Significance and Claims
The authors claim that ProtoPathway provides a complete interpretability chain from individual genes to spatial tissue regions while matching or exceeding the predictive performance of existing methods. The significance lies in its ability to produce "native inference-time attribution" where the model's reasoning is transparent because the representations themselves carry domain-grounded identity. By enforcing stable, task-learned semantic identity on both the morphological and genomic axes, the framework avoids the "black box" nature of post-hoc explanations and the biological incoherence of independent pathway encoders. The paper asserts that this approach enables clinicians to validate model reasoning against established biology, a prerequisite for clinical adoption in high-stakes decision-making. The authors note limitations, including the use of bulk transcriptomics (averaging cell populations) and the fact that attention magnitude does not encode the direction of expression change, requiring separate differential expression analysis for full biological interpretation.