search

Ohyeon5 / syn-search

Optimal Drug Synthesis Condition Searching Conversational AI, using LLMs
MIT License
2 stars 2 forks source link
cheminformatics large-language-models

readme

Syn-Search

Optimal Drug Synthesis Condition Searching Conversational AI using LLMs.

We created an AI framework to help chemists find actionable protocols to synthesize molecules.

Drug discovery often involves meticulous chemical compound synthesis. Medicinal chemists adjust synthesis conditions incrementally to find the optimal state. While using known conditions from literature seems logical, chemical reactions are highly sensitive. Even minor changes in reagents or solvents can alter the outcome significantly.

syn-search

This project aims to create a chat-based search platform using pre-trained LLMs (Hugging Face) for chemists to effortlessly find experimental conditions.

Project Details

Develop a chat-based search platform using pre-trained Language Models (LMs) from hugging face, tailored for chemists to efficiently locate experimental conditions. The platform will leverage the Retrieval Augmented Generation (RAG) approach to cross-reference and bias model responses towards accurate chemistry literature. llama-index will be used to index the chemistry literature. The model is then deployed on a azure static web app. The backend is built with flask and deployed on azure app service.

Chemistry experties information are going to be retrieved from PubChem, Reaxys, ChemArxiv, etc. Then, these dataset will be prepared as vector database for the RAG approach. The conditions to generate the optimal synthesis conditions for a given chemical synthesis name or reaction smart will be curated during the hackathon.

Llama 2

In order to access the Llma 2 models, you need to request access from the Meta website. Requests takes 1-2 days, so please submit the request prior to the usage date.

Currently the codebase uses Azure OpenAI's gpt models, but we will use Llama2 for the future versions.

Project Structure

src
|_ backend
|_ frontend
|_ preprocessing

Setup

Setup python package for model data processing:

make env-data
conda activate syn-data

Setup env for backend:

make env-backend
conda activate syn-ws

Setup env for frontend:

make env-frontend

Dataset

Various chemistry related journals will be used. The dataset will be stored in data folder. Here are some of the journals we are considering:

Reference