data/run_benchmark.py

"Runs model comparison"

from benchmark import visualization
from benchmark import get_cath
from pathlib import Path
import click
import os
import sys


def check_sets(dataset: Path, training_set: Path):
    """Compares training and testing sets, warns if they overlap.
    Parameters
    ----------
    dataset:Path
        Path to .txt file with the dataset.
    training_set:Path
        Path to a file with the training set, can be PISCES, pdb code or pdb+chain.
    """

    with open(training_set) as file:
        training_chains = [x.split()[0][:4].upper() for x in file.readlines()]
        # check for pisces
        if len(training_chains[0]) != 5:
            training_chains = training_chains[1:]
    # check only pdb codes, not chains
    with open(dataset) as file:
        testing_chains = [x.split()[0][:4].upper() for x in file.readlines()]

    repeated_chains = set(testing_chains).intersection(set(training_chains))

    if len(repeated_chains) > 0:
        print(f"{len(repeated_chains)} chains are in both sets:")
        for chain in repeated_chains:
            print(chain)

        print("\n")
        print("Suggested benchmarking set:")
        for chain in testing_chains:
            if chain not in repeated_chains:
                print(chain)
        if click.confirm(
            "Model evaluation might not be valid. Do you want to continue?"
        ):
            click.echo("Continuing!")
        else:
            exit()
    else:
        print("There is no overlap between sets.")


@click.command("compare")
@click.option(
    "--dataset",
    help="Path to .txt file with dataset list (PDB+chain, e.g., 1a2bA).",
    type=click.Path(exists=True),
    required=True,
)
@click.option(
    "--training_set",
    default=False,
    help="Path to .txt file with the training set.",
)
@click.option(
    "--path_to_pdb",
    help="Path to the directory with PDB files.",
    type=click.Path(exists=True),
    required=True,
)
@click.option(
    "--path_to_models",
    help="Path to the directory with .csv prediction files.",
    type=click.Path(exists=True),
    required=True,
)
@click.option(
    "--include",
    help="Path to .txt file with a list of models to be included in comparison. If not provided, 8 models with the best accuracy are compared.",
    type=click.Path(exists=True),
)
@click.option(
    "--torsions",
    is_flag=True,
    help="Produces predicted and true Ramachandran plots for each model.",
)
def compare_models(
    dataset: str,
    path_to_pdb: str,
    path_to_models: str,
    training_set: str,
    include: str = False,
    torsions: bool = False,
) -> None:
    """Generates model summary and comparison plots.
    \f
    Parameters
    ---------
    dataset: str
        Path to .txt file with dataset list (PDB+chain, e.g., 1a2bA).
    path_to_pdb: str
        Path to the directory with PDB files.
    path_to_models: str.
        Path to the directory with .csv prediction files.
    include: str = False
        Path to .txt file with a list of models to be included in comparison. If not provided, 8 models with the best accuracy are compared.
    torsions: bool = False
        Produces predicted and true Ramachandran plots for each model.
    training_set:Path
            Path to a file with the training set, can be PISCES, pdb code or pdb+chain.
    """

    # check training and testing sets
    if training_set:
        check_sets(Path(dataset), Path(training_set))
    else:
        # Warn and ask for confirmation to continue.
        if click.confirm(
            "Cannot compare training and testing sets. YOUR COMPARISON MIGHT NOT BE STATISTICALLY MEANINGFUL. Do you want to continue?"
        ):
            click.echo("Continuing!")
        else:
            exit

    # get model labels to include in comparison
    if include:
        with open(include) as file:
            models_to_include = [x.strip("\n") for x in file.readlines()]
    df = get_cath.read_data(f"{Path(os.path.dirname(sys.argv[0]))/'cath-domain-description-file.txt'}")
    filtered_df = get_cath.filter_with_user_list(df, dataset)
    df_with_sequence = get_cath.append_sequence(
        filtered_df, Path(path_to_pdb)
    )

    accuracy = []
    # load predictions
    list_of_models = {}
    for name in os.listdir(path_to_models):
        if name.split(".")[-1] == "csv":
            path_to_file=Path(path_to_models)/name
            with open(path_to_file.with_suffix('.txt')) as datasetmap:
                model= get_cath.load_prediction_matrix(df_with_sequence, path_to_file.with_suffix('.txt'), path_to_file)
                ignore_uncommon=eval(datasetmap.readline().split()[1])
                pdbs=datasetmap.readline().split()
                if len(pdbs)>1:
                    #visualize accuracy and entropy on pdb files
                    for protein in pdbs[1:]:
                        visualization.show_accuracy(
                            df_with_sequence,
                            protein[:4],
                            model,
                            Path(path_to_models) / f"{name.strip('.csv')}_{protein}.pdb",
                            Path(path_to_pdb),
                            ignore_uncommon,
                        )
            list_of_models[name]=(model,ignore_uncommon)

    for model in list_of_models:
        # make model summary
        visualization.make_model_summary(
            df_with_sequence,
            list_of_models[model][0],
            str(Path(path_to_models) / model),
            Path(path_to_pdb),
            list_of_models[model][1],
        )
        # get overall accuracy
        accuracy.append(
            [
                get_cath.score(
                    df_with_sequence,
                    list_of_models[model][0],
                    ignore_uncommon=list_of_models[model][1],
                )[0][0],
                model,
            ]
        )
        # make Ramachandran plots
        if torsions:
            sequence, prediction, _, angle = get_cath.format_angle_sequence(
                df_with_sequence,
                list_of_models[model][0],
                Path(path_to_pdb),
                ignore_uncommon=list_of_models[model][1],
            )
            visualization.ramachandran_plot(
                sequence,
                list(get_cath.most_likely_sequence(prediction)),
                angle,
                str(Path(path_to_models) / model),
            )
    accuracy = sorted(accuracy)
    # pick 8 best models
    filtered_models = [list_of_models[model[1]][0] for model in accuracy[-8:]]
    ignore_list= [list_of_models[model[1]][1] for model in accuracy[-8:]]
    filtered_labels = [model[1] for model in accuracy[-8:]]
    # include specified models
    if include:
        if len(models_to_include) <= 8:
            for index, model_name in enumerate(models_to_include):
                if model_name not in filtered_labels:
                    filtered_models[index] = list_of_models[model_name][0]
                    ignore_list[index]=list_of_models[model_name][1]
                    filtered_labels[index] = model_name
        else:
            raise ValueError(
                "Too many models are give to plot, select no more than 8 models."
            )

    visualization.compare_model_accuracy(
        df_with_sequence,
        filtered_models,
        filtered_labels,
        Path(path_to_models),
        ignore_list,
    )
    
if __name__=="__main__":
    compare_models()