AbhijeetKumar2125/da3410-ec-cnn

EC CNN Candidate: Enzyme Classification

A character-level residual convolutional neural network for predicting Enzyme Commission (EC) numbers directly from protein amino acid sequences.

Model architecture

• Input: Protein amino acid sequence
• Encoding: Integer encoding of amino acid characters
• Feature extractor: 1D convolutional stem followed by three residual convolutional blocks
• Pooling: Masked mean pooling and masked max pooling
• Classifier: Fully connected neural network
• Output: One predicted EC number per input sequence

Training data

The model was trained on the Leak CURBER enzyme-classification training dataset. Multiple EC annotations for the same sequence were combined during training into a multi-label target.

Inference

The container accepts an input Parquet file containing id and sequence columns and writes JSON Lines predictions.

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

Each output record contains:

{"id": "example_id", "prediction": ["1.1.1.1"]}

The EC number above is an illustrative example, not a prediction for a particular sequence.

Files

• config.json — Model configuration, amino acid vocabulary, and EC class mapping
• model.pt — Trained model weights
• predict.py — Inference implementation
• Dockerfile — Container build configuration

Limitations

This is a candidate model and has not yet been established to outperform the existing retrieval-based submission. Its leaderboard performance must be determined by evaluation.