- Jupyter Notebook 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| result112025 | ||
| .DS_Store | ||
| 0623LLM-based priotization strategy.png | ||
| AD LLM Evaluation.png | ||
| drug_disease_with_numbered_abstracts_20250325_190703.csv | ||
| LLM_validation1125.ipynb | ||
| README.md | ||
| resultanalysis.ipynb | ||
LLM-based Prioritization Framework for Alzheimer’s Therapeutics
This repository contains the LLM-driven evidence synthesis pipeline and the replication guide for the paper: "Bridging the Computational-Experimental Gap: Leveraging Large Language Model to Prioritize Alzheimer’s Therapeutics Based on Comparison of Learning Models".
📂 Repository Contents
-
LLM_validation1125run.ipynb: The master notebook that executes the core pipeline:- Data Loading: Imports the 90 candidate drugs generated by the graph models and the corresponding PubMed abstracts.
- Evidence Synthesis: Queries the GPT-4.1 API.
-
resultanalysis.ipynb: Applies the prioritization criteria to rank the drugs based on the LLM outputs. -
drug_disease_long_format_20250325_190703.csv: Contains the input files.
🚀 How to Run the Code
The workflow is consolidated into the LLM_validation1125run.ipynb notebook for ease of replication.
- Set your OpenAI API key in a file named
openai_api.txt. - Open
LLM_validation1125run.ipynb. - Run all cells to execute the evidence synthesis.
- Run
resultanalysis.ipynbto generate the final lists.
🔬 Replication Guide for Candidate Generation (Graph Models)
The input for LLM_validation1125run.ipynb is a list of candidate drugs. These candidates were generated using three state-of-the-art models. To replicate this upstream step, we utilized the official implementations adapted for the PrimeKG dataset:
1. TxGNN
We utilized the official pre-trained weights and zero-shot prediction scores.
- Repository: https://github.com/mims-harvard/TxGNN
2. CompGCN
We retrained the model on PrimeKG using the standard link prediction setting.
- Repository: https://github.com/malllabiisc/CompGCN
3. RLR (Hetionet)
We computed Degree-Weighted Path Counts (DWPC) using the learn and hetnet_ml packages.