ChemCrow Documentation

By default, ChemCrow uses the RXN4Chem API for retrosynthetic planning and reaction product prediction. To avoid API latency or key requirements, you can self-host these tools using Docker containers.

Run the following commands to start the prediction and retrosynthesis services:

docker run --gpus all -d -p 8051:5000 doncamilom/rxnpred:latest
docker run --gpus all -d -p 8052:5000 doncamilom/retrosynthesis:latest

Before running ChemCrow, you must set your OpenAI API key. You can optionally set a Serp API key for enhanced search capabilities.

export OPENAI_API_KEY=your-openai-api-key

# Optional
export SERP_API_KEY=your-serpapi-api-key

Each tool in ChemCrow is provided as a Docker container that exposes an API for requests. To run a tool, use docker run with the -d flag for detached mode and map a host port to the container's fixed internal port 5000.

For example, to expose the tool on host port 8052, use:

docker run -d -p 8052:5000 doncamilom/rxnpred:latest

Install the ChemCrow package using pip to access reasoning-intensive chemical tools built with Langchain.

pip install chemcrow

To install aizynthfinder along with all its required data and dependencies, use the [all] extra with pip.

This installation method is necessary to ensure that all important files required for running the tool are downloaded and available in your environment.

pip install aizynthfinder[all]

To use the Molecular Transformer tool within the ChemCrow ecosystem, you must manually download the model weights from the provided Google Drive link and place them in the appropriate directory within your setup.

https://drive.google.com/uc?id=1ywJCJHunoPTB5wr6KdZ8aLv7tMFMBHNy

Once a tool container is running, you can send a POST request to the /api/v1/run endpoint. The request body must be a JSON object containing a smiles key with the SMILES string of the reactants.

curl -X POST -H "Content-Type: application/json" -d '{"smiles": "O=C(OC(C)(C)C)c1ccc(C(=O)Nc2ccc(Cl)cc2)cc1"}' http://localhost:8052/api/v1/run

You can interact with the ChemCrow tool APIs using the requests library in Python. The following function sends a SMILES string to the tool's /api/v1/run endpoint and returns the first predicted product from the JSON response.

import json
import requests

def reaction_predict(reactants):
    response = requests.post(
        "http://localhost:8052/api/v1/run",
        headers={"Content-Type": "application/json"},
        data=json.dumps({"smiles": reactants})
    )
    return response.json()['product'][0]

product = reaction_predict('CCOCCCCO.CC(=O)Cl')

To use ChemCrow, import the ChemCrow class from chemcrow.agents. Initialize the agent with a specific model (e.g., gpt-4-0613), temperature, and streaming preference, then use the .run() method to execute chemical queries.

from chemcrow.agents import ChemCrow

chem_model = ChemCrow(model="gpt-4-0613", temp=0.1, streaming=False)
chem_model.run("What is the molecular weight of tylenol?")

When running local Docker containers for RXN4Chem tools, initialize the ChemCrow agent with local_rxn=True to instruct it to use your local services instead of the remote API.

from chemcrow.agents import ChemCrow

chem_model = ChemCrow(model="gpt-4-0613", temp=0.1, streaming=False, local_rxn=True)
chem_model.run("What is the product of the reaction between styrene and dibromine?")