metadata
dataset_info:
features:
- name: Entry
dtype: string
- name: seqs
dtype: string
- name: EC number
dtype: string
- name: Active site
dtype: string
- name: labels
dtype: int64
splits:
- name: train
num_bytes: 50935453
num_examples: 80212
- name: valid
num_bytes: 686688
num_examples: 1000
- name: test
num_bytes: 692134
num_examples: 1000
download_size: 49792005
dataset_size: 52314275
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: valid
path: data/valid-*
- split: test
path: data/test-*
Uniprot (ec:) AND (go_manual:) on 5/30/25
Remove any EC entries with '-'
Remove any EC entries with more than one EC number
Filter for sequences between 20 and 2048
Cluster at 80% sequence similarity
docker run -v pwd:/data -w /data cd-hit cd-hit -i ec.fasta -o output_ec_80 -c 0.8 -n 3 -T 72 -M 512000
Keep representative sequences
Keep EC numbers with 100 or more examples
Map each EC number to a unique integer
Remove any duplicates, favor entries with active site information
Shuffle
Split into 1000 valid and 1000 test randomly, make sure they all have active site info
13.8% of train entries have active site info