# ASM-HHD

**Repository Path**: Tomhappy/ASM-HHD

## Basic Information

- **Project Name**: ASM-HHD
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-01-11
- **Last Updated**: 2026-01-11

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# ASM-HHD 

![architecture](ASMHHD.jpg)
ASM-HHD is an implementation of the 
[Adaptive Multi-Scale Hypergraph Heat Diffusion for Interpretable ncRNA–Disease Regulatory Modeling] 


Official Minimal Working Implementation (PyTorch)

This repository provides the clean and reproducible implementation of our model ASM-HHD, as described in the manuscript:

Adaptive Spectral Multi-Scale Hypergraph Heat Diffusion for ncRNA–Disease Regulatory Modeling

The released code keeps all algorithmic components exactly consistent with the paper, while removing visualization, comments, and redundant logic to produce a concise, ready-to-run version.

🔥 Highlights

Dual-channel spectral diffusion

Global channel: exact heat kernel 

Local channel: learnable Gegenbauer-based heat diffusion

Multi-modal alignment


Noise-adaptive missing-token mechanism

Hypergraph learning

High-order ncRNA regulatory structures modeled by normalized hypergraph Laplacian

Interpretable

Diffusion times, spectral basis, and fusion weights are fully learnable

Complete training + testing pipeline

Produces results/TestPredictions.csv including real node IDs

📁 ASM-HHD  
├── 🚀 main.py  
├── 🧠 train.py  
├── 🔧 model.py  
├── 🌐 globalC.py  
├── 📡 localC.py  
├── 📥 loader.py  
├── 🛠 utility.py  
├── 📊 metric.py  
├── 🔗 pretrain_align.py  
└── 📂 results/  


All functions strictly match those described in the “Method” section of the manuscript.

🧬 Method Overview (Paper Alignment)
1. Multimodal Feature Alignment (MA-FAR)

pretrain_align.py

Encodes RNA sequences via RiNALMo (1280-dim)

Encodes disease MeSH text via BioBERT (768-dim)

Shared latent space projection (64-dim)

Missing-token replacement + node-wise sinusoidal noise

Outputs aligned features:

aligned_features/all_nodes_aligned_64d.csv

2. Hypergraph Construction

loader.py

Loads CDAtrain.csv / CDAtest.csv

Maps real node IDs

Builds high-order hyperedges

Train/val/test split with reproducible seeds

3. Multi-Scale Spectral Hypergraph Diffusion
✔ Global Channel — Exact Heat Kernel

globalC.py

Captures long-range, cross-module propagation

Learnable global diffusion time 

✔ Local Channel — Learnable Gegenbauer Kernel

localC.py

Node-wise diffusion time 

Learnable Gegenbauer λ controlling high/low-frequency smoothing

Captures localized structural heterogeneity

✔ Fusion

model.py uses multi-head attention:


Produces unified node embeddings used for link prediction.

4. Training & Evaluation

train.py implements:

Task loss (binary cross-entropy)

Global–local contrastive consistency

Spectral smoothness regularization

Diffusion-time regularization

Saves:

results/TestPredictions.csv


(with real node IDs, scores, labels)

🚀 How to Run
1. Prepare aligned features
python pretrain_align.py

2. Train the model
python main.py \
    --data CDAtrain.csv \
    --feature_dir aligned_features \
    --model dual \
    --epochs 200 \
    --val_mode 1

3. Test results

Automatically generated in:

results/TestPredictions.csv

📊 Datasets Used

circRNA–miRNA–cancer:
http://www.biobdlab.cn:8000

lncRNA–miRNA–cancer:
http://www.rnanut.net/lncrnadisease

miRNA–target gene:
https://mirtarbase.cuhk.edu.cn/∼miRTarBase/miRTarBase_2025

CMI-9589 / CMI-9905:
https://github.com/1axin/KGDCMI

circad & circ2Traits:
https://github.com/lanbiolab/CompareExperiment

miRNA–disease (HMDD):
http://www.cuilab.cn/hmdd

curated in https://github.com/pingxuan-hlju/SFPred

🧪 Environment
Python 3.8–3.10
PyTorch ≥ 1.10
numpy
pandas
DHG ≥ 0.9
scikit-learn ≥ 1.0

📚 Citation

If you use this code:

[]

🎯 Notes

This repository includes all algorithmic components described in the paper

All visualization components have been removed

The code is optimized for clarity, reproducibility, and minimal dependencies

globalC/localC naming matches the paper’s global/local spectral diffusion channels






## 📄 License

This project is released under the **MIT License**, which allows reuse, modification, and distribution for both academic and commercial purposes, provided that the original copyright notice is preserved.