README.md

RetroSynthesisAgent

code for paper "Leveraging Large Language Models as Knowledge-Driven Agents for Reliable Retrosynthesis Planning".

This project aims to conduct retrosynthesis planning for polymer materials based on literature, utilizing an LLM agent and knowledge graphs.

Requirements

conda create -n retrosyn python=3.11
conda activate retrosyn
pip install rdkit requests python-dotenv PyMuPDF scholarly openai networkx graphviz pubchempy Pillow fastapi pydantic uvicorn pyvis loguru

Data

• eMolecules download URL: https://downloads.emolecules.com/free/2024-07-01/
• Download it and turn it to a set as a format of json

env. File Setting

• Set your OpenAI API key (API_KEY) and optional BASE_URL to use the LLM.
• Set the HEADERS and COOKIES from your browser for web scraping of literatures in JSON
• an example:

API_KEY=xxx
BASE_URL=xxx
HEADERS={"user-agent": "xxx"}
COOKIES={"xxx": "xxx", "xxx": "xxx"}

Run the demo

sh runRetroSynAgent.sh

Parameter Description:

• --material: Specifies the material to be processed.
• --num_results: Defines the number of PDF to be processed.
• --alignment: Determines whether to align entities.
• --expansion: Determines whether to expand the tree with additional literature.
• --filtration: Determines whether to apply filtration to the reactions.

We provide a demo video of its automated operation process:

• After constructing the chemical retrosynthetic pathway tree for the target substance, a URL will be generated. You can open it directly in your local browser.
• When you hover over a node in the tree, the name of the substance represented by that node will be displayed.