This repository contains an implementation of the HINN (Hierarchical Integration Neural Network) model, designed to learn complex relationships across multi-omic data layers for predicting cognitive scores (e.g., MMSE). The architecture supports modular interpretation through DeepLIFT and visualizes pathway influence using Sankey diagrams.
The HINN model integrates the following omic layers:
- SNPs
- Methylation Sites
- Gene Expression
- Demographics
Each layer is connected through biologically-informed sparse matrices:
- SNP-to-Methylation
- Methylation-to-Gene
- Gene-to-Pathway
Custom masked transformation using the sparse SNP–Methylation connectivity matrix.
Identity mapping with learnable weights constrained by an identity mask.
SNP-based transformation and methylation data are combined via element-wise multiplication followed by a non-linear transformation.
- Output is concatenated with a dense-transformed SNP layer.
- Further processed through Primary/Secondary layers for methylation → expression.
- Division and nonlinear operations refine the gene layer.
- Pathway transformation is appended.
- Dense layers with batch normalization and dropout extract predictive signals.
- Demographic data is fused near the output layer.
- Captum DeepLIFT Integration – Interpretation of feature importance.
- Plotly Sankey Visualization – Visual flow of top omic features to pathway level.
- Custom Layers – Enforces biological priors through sparse matrices.
HINN_model_deep_lift.py– Main training and interpretation script.*.csv– Input files: omics datasets and sparse connectivity matrices.requirements.txt– Python package dependencies.
python HINN_model_deep_lift.py