RedEx

Predicting experimental redox potentials for aqueous organic molecules using GNNs with transfer learning.

RedEx is a GAT ensemble, pre-trained on computational DFT-calculated redox potentials and fine-tuned on experimental redox potentials curated from the literature.

Installation

conda env create -f environment.yml
conda activate redex
pip install -e .

Pre-training

Pre-train on the computational RedoxDB dataset:

python scripts/pretrain.py --model gat        # single model
python scripts/pretrain.py --all              # all five architectures

Configs are in configs/pretrain_{model}.yaml. Checkpoints are saved to weights/pretrained/.

Fine-tuning

Fine-tune on experimental data:

python scripts/finetune.py \
    --model gat \
    --pretrained weights/pretrained/gat_full_pretrain.pth \
    --config configs/finetune_gat.yaml

Prediction

Pre-trained and fine-tuned weights are included and you can use RedEx to predict redox potentials with the below code.

python scripts/predict.py \
    --model-dir weights/finetuned/gat_partial \
    --smiles "OC1=CC=CC2=NC3=CC=CC=C3N=C21"

Or from a CSV (needs an oxidised_smiles column):

python scripts/predict.py \
    --model-dir weights/finetuned/gat_partial \
    --csv data/Experimental_redox_potentials.csv \
    --output predictions.csv

How to cite

To be added upon publication.

Licence

MIT