Reaction Catalyst Prediction

This project predicts the catalyst required for a given chemical reaction. It uses graph-based ML (Morgan fingerprints via RDKit) trained on the ORDerly dataset and exposes a FastAPI + Frontend for easy prediction.

---

Quick Start

Clone the repo

git clone https://github.com/dshryn/graph-catalyst-ml.git
cd graph-catalyst-ml

Create environment

conda create -n catalyst-ml python=3.10 -y
conda activate catalyst-ml
conda install -c conda-forge rdkit tqdm scikit-learn fastapi uvicorn pandas numpy joblib -y

Dataset

Download the ORDerly dataset from Figshare: https://figshare.com/articles/dataset/ORDerly_chemical_reactions_condition_benchmarks/23298467?file=44413037

Place it like:

graph-catalyst-ml/data/
 ├─ orderly_condition_with_rare_train.csv
 └─ orderly_condition_with_rare_test.csv

Preprocessing Pipeline

Clean RXN SMILES

python scripts/clean_rxn_str.py --input ./data/orderly_condition_with_rare_train.csv --output ./data/tmp_train_cleaned.csv --chunksize 2000 --max-rows 5000 --force-overwrite

Normalize + Explode Agents

python scripts/normalize_and_explode_agents.py --input ./data/tmp_train_cleaned.csv --output ./data/orderly_condition_with_rare_train_normalized.csv

Skip malformed lines

python scripts/skip_bad_lines.py

This creates:

data/orderly_condition_with_rare_train_norm_whitelist_skipbad.csv

Balance dataset

python scripts/sample_balance.py --input ./data/orderly_condition_with_rare_train_norm_whitelist_skipbad.csv --output ./data/orderly_condition_with_rare_train_balanced.csv --max-samples 2000

Training

Run:

python -m src.train --train-csv "./data/orderly_condition_with_rare_train_balanced.csv" --test-csv "./data/orderly_condition_with_rare_test_norm_whitelist.csv" --out "./models/artifacts_balanced.joblib" --nbits 128 --chunk-size 500 --max-rows 30000 --nn-sample 1000 --exclude-other --class-weight balanced --random-state 42

Model artifact:

models/artifacts_balanced.joblib

Debug / Local Inference

python scripts/debug_inference.py --artifact ./models/artifacts_balanced.joblib --smiles "CC(=O)c1cc(C)cc([N+](=O)[O-])c1O>>CC(=O)c1cc(C)cc(N)c1O"

Expected Fe / Cu / Pd (for nitro reduction reactions)

Run API + Frontend

Start API

uvicorn src.api_server:app --reload --port 8000

Open Frontend

Open web/index.html in a browser and enter required fields to predict the catalyst.