# From Black Box to Biological Insight: AttentioFuse Unlocks Multi-Omics Dynamics in Lung Cancer

## TL;DR

This paper introduces AttentioFuse, an interpretable deep learning model that improves lung cancer staging by integrating multi-omics data and providing biological insights.

## Contribution

The novelty lies in the Reactome-guided mid-fusion framework that links predictions to gene-pathway evidence chains, enhancing interpretability in multi-omics models.

## Key findings

- AttentioFuse achieves competitive performance in T- and N-stage prediction for lung cancer.
- The model reveals biologically relevant pathways like AKT/mTOR and Notch signaling in lung cancer subtypes.
- It provides pathway-resolved explanations that are consistent across different data folds.

## Abstract

Lung cancer is often staged using imaging and pathology, but tumors within the same stage can behave very differently. Multi-omics profiles contain complementary information that may improve risk stratification, yet many deep learning models act as “black boxes” that are hard to audit clinically. We present AttentioFuse, an interpretable Reactome-guided mid-fusion framework that links predictive signals to a gene–pathway–modality evidence chain. In TCGA LUAD/LUSC cohorts, AttentioFuse achieves competitive performance for T- and N-stage prediction and provides pathway-resolved explanations that are coherent across folds. Importantly, these explanations are hypothesis-generating and do not imply causality without experimental validation.

Background: Lung adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC), the major subtypes of non-small cell lung cancer (NSCLC), exhibit distinct molecular landscapes that demand precision in prognosis and therapy. While deep learning models can achieve high predictive accuracy, their black-box nature limits clinical translation. Methods: We introduce AttentioFuse, an interpretable deep learning framework employing a Reactome-guided mid-fusion strategy for multi-omics integration. AttentioFuse builds on three pillars: (i) dual-phase learning with omics-specific encoders to preserve modality-unique patterns, (ii) hierarchical attention mechanisms (cross-omics, feature-level, and fusion-layer) to quantify layer contributions dynamically, and (iii) integrated explainability combining DeepSHAP and global attention weights for gene-to-pathway interpretation. Two depth variants are instantiated under identical priors: a three-layer configuration (3F) for main discrimination and a five-layer configuration (AttentioFuse-5X) for deeper hierarchical interpretation; the 5X variant is trained end-to-end and yields comparable accuracy while enhancing pathway-level resolution. Results: Evaluated on The Cancer Genome Atlas (TCGA) LUAD/LUSC cohorts, AttentioFuse matches state-of-the-art performance in TNM staging while uncovering actionable biological insights, including pan-NSCLC AKT/mTOR metabolic control, histology-divergent Notch signaling roles, and additional pathways related to developmental reactivation, microbiota-associated metastasis, and extracellular matrix remodeling. Conclusions: By design, AttentioFuse-5X bridges predictive performance with hierarchical, pathway-resolved explanations, advancing oncology by transforming black-box predictions into biologically grounded decision support.