DuSSM: Disentangling Cognitive Duality in Biomedical Linguistics

This repository contains the official implementation of DuSSM, a parallel State-Space framework designed to resolve the "feature entanglement" in biomedical relation extraction. By structurally isolating localized morphological triggers (Explicit Stream) from long-range semantic logic (Implicit Stream), DuSSM achieves superior clinical reliability and $O(N)$ inference efficiency.

🌟 Key Contributions

• Parallel Cognitive Disentanglement: Structurally separates morphological feature extraction from topological reasoning to resolve feature entanglement.
• Selective State-Space Formulation: Models biomedical text as a continuous dynamic system using the Mamba architecture, effectively capturing long-range causal dependencies.
• Clinical Determinism: Establishes a "determinism boundary" with 94.32% precision on non-interaction cases, significantly mitigating "alert fatigue" in pharmacovigilance.
• Computational Efficiency: Leverages linear-time complexity $O(N)$ for a 1.6x inference speedup over Transformer baselines.

📊 Benchmark Results

DuSSM has been rigorously validated across four heterogeneous biomedical datasets:

Dataset	Type	Metric	Result
DDI-2013	Drug-Drug Interaction	Macro F1	82.27%
ChemProt	Chemical-Protein	Micro F1	87.64%
GAD	Gene-Disease	F1/AUC	83.55%
EU-ADR	Gene-Disease	F1/AUC	86.04%

⚙️ Installation

# Clone the repository
git clone [https://github.com/Hero-Legend/DuSSM.git](https://github.com/Hero-Legend/DuSSM.git)
cd DuSSM

# Create environment
conda create -n dussm python=3.10
conda activate dussm

# Install core dependencies
pip install torch==2.4.0 mamba-ssm causal-conv1d transformers

🚀 Usage

1. Data Preparation

The experimental scripts support DDI-2013, ChemProt, GAD, and EU-ADR. Place your preprocessed datasets in the data/ directory.

2. Training & Evaluation

To run the primary DDI extraction task:

python CNN+transformer-mamba.py --dataset ddi2013 --batch_size 32 --lr 1e-4 --epochs 22

To run gene-disease association experiments:

python GAD_expriment.py --fold 10
python EUADR_expriment.py --fold 10

🔍 Mechanistic Insight

DuSSM utilizes Cognitive Gating to filter linguistic noise. The selection intensity ($\Delta_t$) adaptively contracts to skip clinical jargon ($\downarrow$ 62%) and expands to record critical relational anchors.

📄 License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.