Legend7001/ec-cnn-scratch

Enzyme classification: 1-D CNN trained from scratch

DA3410 project submission for the Leak CURBER enzyme classification task. This is a comparison approach, no outside data and no similarity search.

1. Each letter of the sequence becomes a number (20 amino acids, 1 for anything else); sequences are cut at 1500 letters.
2. A 1-D convolutional network (an embedding, 5 convolution layers with window 9 and dilations 1, 2, 4, 8, max and mean pooling over the sequence, one linear layer) gives one score per EC that occurs in training (4614).
3. The EC with the highest score is the answer.

Trained on all training sequences: binary cross-entropy, AdamW (lr 2e-3, weight decay 1e-2), one-cycle learning rate, batch normalisation, dropout 0.2, 20 epochs.

• predict.py: runs the network on CPU. ecnn.py: the network. weights/cnn.pt: the trained weights and the EC list.
• inputs.py, task.json, ec_vocabulary.json: from the course starter. weights/baseline.json: fallback EC.