GleghornLab/enzyme_kcat

DLKcat (BRENDA and SABIO-RK) with splits from Biomap, and repeated and short sequences removed. Enzymes with multiple reactions have their kcat averaged.

The kcat is log10 normalized, so the unit is log10(1/s). However, because it is averaged over reactions and also reaction ambiguous, it is really just a general proxy for catalytic rate. Higher is faster.

Processing:

import pandas as pd
from datasets import Dataset, DatasetDict, concatenate_datasets

def process_dataset(dataset_dict):
    precedence = ['train', 'valid', 'test']
    # Add a 'split' column to each dataset
    for split in dataset_dict.keys():
        dataset_dict[split] = dataset_dict[split].add_column('split', [split]*len(dataset_dict[split]))
    # Concatenate all splits into one dataset
    all_data = concatenate_datasets([dataset_dict[split] for split in dataset_dict.keys()])
    # Convert to pandas DataFrame
    df = all_data.to_pandas()
    # Remove sequences with length less than 50
    df['seq_length'] = df['seqs'].apply(len)
    df = df[df['seq_length'] >= 50]
    # Group by 'seqs' to find duplicates and average the labels
    def aggregate_group(group):
        avg_label = group['labels'].mean()
        # Assign the sequence to the highest-precedence split it appears in
        for p in precedence:
            if p in group['split'].values:
                selected_split = p
                break
        return pd.Series({'labels': avg_label, 'split': selected_split})
    df_grouped = df.groupby('seqs').apply(aggregate_group).reset_index()
    # Split the DataFrame back into the original splits without overlapping sequences
    new_dataset_dict = DatasetDict()
    for split in precedence:
        df_split = df_grouped[df_grouped['split'] == split]
        new_dataset_dict[split] = Dataset.from_pandas(df_split[['seqs', 'labels']], preserve_index=False)
    return new_dataset_dict

From DLKcat paper