krishdange27/enzyme-retrieval-da3410

DA3410 Enzyme Retrieval — Soft-target Morgan MLP

CPU inference package for the Reaction SMILES → EC prediction task.

Contents

• predict.py: competition inference entrypoint (--input, --output)
• weights/final_mlp_softce.pt: trained checkpoint
• ec_vocabulary.json: official EC vocabulary in model class order
• Dockerfile: Linux x86-64 CPU image
• requirements.txt: runtime dependencies

Inference

The evaluator invokes:

python /app/predict.py --input INPUT.parquet --output OUTPUT.jsonl

The input Parquet must contain id and rxn_smiles. Output is JSONL with exactly id and prediction; each prediction contains one distinct valid EC. Identical reaction SMILES are grouped, retaining the first ID, as required by the task.

Local Docker test

docker build --platform linux/amd64 -t enzyme-retrieval .
mkdir -p output
docker run --rm --platform linux/amd64 --network none --cpus 2 --memory 16g \
  -v "$PWD/../data:/input:ro" -v "$PWD/output:/output" \
  enzyme-retrieval python /app/predict.py \
  --input /input/val_inputs.parquet --output /output/predictions.jsonl

Important

The checkpoint and vocabulary must be copied from the completed Colab training run. Do not use a checkpoint from a different experiment. The inference code reads the fingerprint configuration, hidden layer sizes, log1p setting, and EC order from the checkpoint and reproduces the notebook's Morgan feature pipeline.