Understanding disease mechanisms requires modeling the coordinated activity of multiple biological pathways rather than treating them as independent entities. Most existing pathway-based deep learning approaches aggregate pathway representations using fully connected layers, ignoring interactions between pathways. This limitation can restrict both predictive performance and biological interpretability.
Disease progression is driven by complex interactions across biological pathways. However, current models:
- Treat pathways independently
- Ignore inter-pathway relationships
- Miss biologically meaningful interactions
To address this gap, we propose a model that explicitly learns pathway–pathway interactions.
We introduce PINT (Pathway–Pathway Interaction Network), a deep learning framework that:
- Learns pathway representations from gene expression data
- Models interactions between pathways using a self-attention mechanism
- Uses an attention-based pooling layer to identify patient-specific pathway importance
We evaluated PINT on five TCGA cancer datasets for survival analysis.
Key findings:
- Consistently outperforms benchmark models
- Identifies pathways significantly associated with survival
- Reveals biologically meaningful pathway–pathway interactions
On the BRCA dataset:
- PINT identified important survival-associated pathways
- Discovered patient-specific pathway interactions
- Many findings are supported by existing literature
Notably:
- The RAS signaling pathway was strongly associated with survival
- Learned interactions captured known relationships with:
- cAMP signaling
- TNF signaling
- Rap1 signaling
- Introduces a novel framework for modeling pathway–pathway interactions
- Improves both predictive performance and interpretability
- Provides patient-specific biological insights
data_preprocess/ # Contains TCGA-BRCA dataset and preprocessing scripts
exp_setup/ # Train/validation/test split generation
exp_runs/ # Model training scripts (Slurm + Python)
metadata/ # KEGG pathway information