Attention based deep learning model to classify a given protein sequence into the seven classes of enzymes or a negative class. The code for the machine learning and deep learning models (with cross validation and sampling methods), along with the code for preprocessing (vectorization) and postprocessing (statistical testing and plotting) can be found in the /src folder. Run the machine learning models by running python3 ml_models.py in the directory. DL models can be run using the script run_dl_models.py which can be run by a command of the format:

python3 run_dl_models.py CNN --epochs 200 --batch 256

epochs and batch are optional arguments which default to 100 and 128 respectively, if not specified. The model used has to be the first argument to the script and has to be one of CNN, LSTM, BILSTM, GRU, ABLE.

Data used is provided in the data.zip file (64MB) - containing two pickle files (X.pickle and y.pickle), which contain the vectorized representations of the data. X.pickle contains an array of shape (127537, 3, 100) - 1,27,537 proteins represented in vectorized form of size (3, 100). Y.pickle contains the labels for these data (numbers 0 to 7), with 0 representing the negative class. Before running any of the code in src/, make sure to extract data.zip in the root of this repository to create the data/ directory with these two files.

The results directory contains all information about the performance of the models - including runtimes, multiclass confusion matrices, f-score, precision, and recall, among other metrics, as both Python pickle files and CSVs.

• results/dl contains the training history for each run of the models, stored as .npy files - the syntax for each filename in this directory is {{MODEL_NAME}}_{{SAMPLING_METHOD}}_{{TESTING_FOLD}}_{{NUM_EPOCHS}}_{{BATCH_SIZE}}.npy.
• Performance metrics of the runs are stored in these pickle files: ABLE_results.pickle BILSTM_results.pickle CNN_results.pickle GRU_results.pickle LSTM_results.pickle ML_results.pickle, in the results directory
• The Jupyter Notebook in results/postprocessing contains the scripts to save Wilcoxin Signed Rank Test results between all pairs of models to files, on all performance metrics (precision, recall, f-score, balanced accuracy).

If you find these code or results useful in your research, please consider citing:

@article{ABLE2021,
    title = {ABLE: Attention Based Learning for Enzyme Classification},
    journal = {Computational Biology and Chemistry},
    pages = {107558},
    year = {2021},
    issn = {1476-9271},
    doi = {https://doi.org/10.1016/j.compbiolchem.2021.107558},
    url = {https://www.sciencedirect.com/science/article/pii/S1476927121001250},
    author = {Nallapareddy Mohan Vamsi and {Rohit Dwivedula}},
}

A preprint of this article is also available on BioRxiv, though the preprint does not contain results from ADASYN sampling and comparisions with DeepEC (these additions were made during the peer review process).