Statistical Modelling

Calculates differences in the probability distributions of each feature for the different classes. This is only available to supervised datasets (i.e. data must has class labels).

2 Classes for end user usage:

1. ClassificationStatModel For analysis when the target data is categorical (classification). Can only use two classes at a time (i.e., binary classification)

2. RegressionStatModel For analysis when the target data is continuous (regression).

These classes both inherit from a parent class called "_ProteinStatModel" which abstracts as much as their shared behavior as possible.

ClassificationStatModel dataclass

Bases: _ProteinStatModel

Handles the generation of statistical models for PyContact data sets when the target is made up of two unqiue class labels.

Note that most attributes are inherited from _ProteinStatModel.

Attributes

pd.DataFrame

Input dataframe.

list

Class labels inside the column "Target" of the dataset to model. You can only use two classes for this approach.

str

Directory path to store results files to. Default = ""

list, optional

What types of molecular interactions to generate the correlation matrix for. options are one or more of: ["Hbond", "Hydrophobic", "Saltbr", "Other"] Default is to include all 4 types.

pd.DataFrame

Input dataset with all features scaled.

list

List of all feature labels in the dataset.

np.ndarray

Values on the x-axis for plotting the kernel density estimations.

dict

Nested dictionary. Outer layer keys are class names, and values are a dictionary of each feature (as inner key) and values of a nested array of kernel density estimations (kdes).

dict

Dictionary with each feature's (keys) and Jensen Shannon distance (values). Dictionary is sorted from largest Jensen Shannon distance to smallest.

dict

Dictionary with each feature's (keys) and mutual informations (values). Dictionary is sorted from largest mutual information to smallest.

Methods

calc_mutual_info_to_target(save_result=True) Calculate the mutual information between each feature to the target classes.

calc_js_distances(kde_bandwidth=0.02, save_result=True) Calculate the Jensen-Shannon (JS) distance (metric) between each feature to the target classes.

Source code in key_interactions_finder/stat_modelling.py

@dataclass
class ClassificationStatModel(_ProteinStatModel):
    """
    Handles the generation of statistical models for PyContact data sets
    when the target is made up of two unqiue class labels.

    Note that most attributes are inherited from _ProteinStatModel.

    Attributes
    ----------

    dataset : pd.DataFrame
        Input dataframe.

    class_names : list
        Class labels inside the column "Target" of the dataset to model.
        You can only use two classes for this approach.

    out_dir : str
        Directory path to store results files to.
        Default = ""

    interaction_types_included : list, optional
        What types of molecular interactions to generate the correlation matrix for.
        options are one or more of: ["Hbond", "Hydrophobic", "Saltbr", "Other"]
        Default is to include all 4 types.

    scaled_dataset : pd.DataFrame
        Input dataset with all features scaled.

    feature_list : list
        List of all feature labels in the dataset.

    x_values : np.ndarray
        Values on the x-axis for plotting the kernel density estimations.

    kdes : dict
        Nested dictionary. Outer layer keys are class names, and values
        are a dictionary of each feature (as inner key) and values of a
        nested array of kernel density estimations (kdes).

    js_distances : dict
        Dictionary with each feature's (keys) and Jensen Shannon distance (values).
        Dictionary is sorted from largest Jensen Shannon distance to smallest.

    mutual_infos : dict
        Dictionary with each feature's (keys) and mutual informations (values).
        Dictionary is sorted from largest mutual information to smallest.

    Methods
    -------

    calc_mutual_info_to_target(save_result=True)
        Calculate the mutual information between each feature to the target classes.

    calc_js_distances(kde_bandwidth=0.02, save_result=True)
        Calculate the Jensen-Shannon (JS) distance (metric) between each feature to
        the target classes.
    """

    # Generated at initialization.
    class_names: list = field(default_factory=[])
    # Generated later.
    x_values: np.ndarray = field(init=False)
    kdes: dict = field(init=False)
    js_distances: dict = field(init=False)

    # Called at the end of the dataclass's initialization procedure.
    def __post_init__(self) -> None:
        """Filter, rescale and calc the kdes for each feature."""
        self.out_dir = _prep_out_dir(self.out_dir)

        self.x_values = np.empty(shape=(0, 0))
        self.kdes = {}
        self.js_distances = {}
        self.mutual_infos = {}

        if sorted(self.interaction_types_included) != sorted(
                ["Hbond", "Hydrophobic", "Saltbr", "Other"]):
            self.dataset = _filter_features_by_strings(
                dataset=self.dataset,
                strings_to_preserve=self.interaction_types_included
            )

        self.feature_list = list(self.dataset.columns)
        self.feature_list.remove("Target")

        # Features need to be scaled in order to use same bandwidth throughout.
        self.scaled_dataset = self._scale_features()

        if len(self.class_names) != 2:
            raise ValueError(
                "The number of classes to compare should be 2. \n" +
                "Please use a list of 2 items for the parameter: 'class_names'.")

    def calc_mutual_info_to_target(self, save_result: bool = True):
        """
        Calculate the mutual information between each feature to the 2 target classes.
        Note that Sklearns implementation (used here) is designed for "raw datasets"
        (i.e., do not feed in a probability distribution, instead feed in the observations).

        Further, the mutual information values calculated from Sklearns implementation are
        scaled by the natural logarithm of 2. In this implementation,
        the results are re-scaled to be linear.

        Parameters
        ----------

        save_result : Optional[bool] = True
            Save result to disk or not.
            Optional, default is to save.
        """
        df_features = self.scaled_dataset.drop("Target", axis=1)
        features_array = df_features.to_numpy()
        classes = self.scaled_dataset["Target"].to_numpy()

        mutual_info_raw = mutual_info_classif(features_array, classes)
        mutual_info_rescaled = np.around((np.exp(mutual_info_raw) - 1), 5)

        self.mutual_infos = dict(
            zip(df_features.columns, mutual_info_rescaled))
        self.mutual_infos = {k: v for k, v in sorted(
            self.mutual_infos.items(), key=lambda item: item[1], reverse=True)}

        print("Mutual information scores calculated.")

        if save_result:
            out_file_path = Path(
                self.out_dir, "Mutual_Information_Per_Feature_Scores.csv")
            self._per_feature_scores_to_file(
                per_feat_values=self.mutual_infos,
                out_file=out_file_path
            )
            print(
                "You can also access these results via the class attribute: 'mutual_infos'.")

    def calc_js_distances(self, kde_bandwidth: float = 0.02, save_result: bool = True):
        """
        Calculate the Jensen-Shannon (JS) distance (metric) between each feature to
        the target classes.
        Requires that each feature is described by a probabilty distribution.

        Parameters
        ----------

        kde_bandwidth : Optional[float]
            Bandwidth used to generate the probabilty distribtions for each feature set.
            Note that features are all scaled to be between 0 and 1 before this step.
            Optional, default = 0.02

        save_result : Optional[bool] = True
            Save result to disk or not.
            Optional, default is to save.
        """
        for class_name in self.class_names:
            # split observations into each class first.
            per_class_dataset = self.scaled_dataset[(
                self.scaled_dataset["Target"] == class_name)]

            # generate kdes for each class.
            self.x_values, self.kdes[class_name] = self._gen_kdes(
                input_features=per_class_dataset,
                kde_bandwidth=kde_bandwidth)

        # iterate through each feature and calc js dist.
        for feature in self.feature_list:
            distrib_1 = self.kdes[self.class_names[0]][feature]
            distrib_2 = self.kdes[self.class_names[1]][feature]

            js_dist = np.around(jensenshannon(distrib_1, distrib_2, base=2), 5)

            self.js_distances.update({feature: js_dist})

        self.js_distances = {k: v for k, v in sorted(
            self.js_distances.items(), key=lambda item: item[1], reverse=True)}

        print("Jensen-Shannon (JS) distances calculated.")

        if save_result:
            out_file_path = Path(
                self.out_dir, "Jensen_Shannon_Per_Feature_Scores.csv")
            self._per_feature_scores_to_file(
                per_feat_values=self.js_distances,
                out_file=out_file_path
            )

            print(
                "You can also access these results via the class attribute: 'js_distances'.")

__post_init__()

Filter, rescale and calc the kdes for each feature.

Source code in key_interactions_finder/stat_modelling.py

def __post_init__(self) -> None:
    """Filter, rescale and calc the kdes for each feature."""
    self.out_dir = _prep_out_dir(self.out_dir)

    self.x_values = np.empty(shape=(0, 0))
    self.kdes = {}
    self.js_distances = {}
    self.mutual_infos = {}

    if sorted(self.interaction_types_included) != sorted(
            ["Hbond", "Hydrophobic", "Saltbr", "Other"]):
        self.dataset = _filter_features_by_strings(
            dataset=self.dataset,
            strings_to_preserve=self.interaction_types_included
        )

    self.feature_list = list(self.dataset.columns)
    self.feature_list.remove("Target")

    # Features need to be scaled in order to use same bandwidth throughout.
    self.scaled_dataset = self._scale_features()

    if len(self.class_names) != 2:
        raise ValueError(
            "The number of classes to compare should be 2. \n" +
            "Please use a list of 2 items for the parameter: 'class_names'.")

calc_js_distances(kde_bandwidth=0.02, save_result=True)

Calculate the Jensen-Shannon (JS) distance (metric) between each feature to the target classes. Requires that each feature is described by a probabilty distribution.

Parameters

Optional[float]

Bandwidth used to generate the probabilty distribtions for each feature set. Note that features are all scaled to be between 0 and 1 before this step. Optional, default = 0.02

Optional[bool] = True

Save result to disk or not. Optional, default is to save.

Source code in key_interactions_finder/stat_modelling.py

def calc_js_distances(self, kde_bandwidth: float = 0.02, save_result: bool = True):
    """
    Calculate the Jensen-Shannon (JS) distance (metric) between each feature to
    the target classes.
    Requires that each feature is described by a probabilty distribution.

    Parameters
    ----------

    kde_bandwidth : Optional[float]
        Bandwidth used to generate the probabilty distribtions for each feature set.
        Note that features are all scaled to be between 0 and 1 before this step.
        Optional, default = 0.02

    save_result : Optional[bool] = True
        Save result to disk or not.
        Optional, default is to save.
    """
    for class_name in self.class_names:
        # split observations into each class first.
        per_class_dataset = self.scaled_dataset[(
            self.scaled_dataset["Target"] == class_name)]

        # generate kdes for each class.
        self.x_values, self.kdes[class_name] = self._gen_kdes(
            input_features=per_class_dataset,
            kde_bandwidth=kde_bandwidth)

    # iterate through each feature and calc js dist.
    for feature in self.feature_list:
        distrib_1 = self.kdes[self.class_names[0]][feature]
        distrib_2 = self.kdes[self.class_names[1]][feature]

        js_dist = np.around(jensenshannon(distrib_1, distrib_2, base=2), 5)

        self.js_distances.update({feature: js_dist})

    self.js_distances = {k: v for k, v in sorted(
        self.js_distances.items(), key=lambda item: item[1], reverse=True)}

    print("Jensen-Shannon (JS) distances calculated.")

    if save_result:
        out_file_path = Path(
            self.out_dir, "Jensen_Shannon_Per_Feature_Scores.csv")
        self._per_feature_scores_to_file(
            per_feat_values=self.js_distances,
            out_file=out_file_path
        )

        print(
            "You can also access these results via the class attribute: 'js_distances'.")

calc_mutual_info_to_target(save_result=True)

Calculate the mutual information between each feature to the 2 target classes. Note that Sklearns implementation (used here) is designed for "raw datasets" (i.e., do not feed in a probability distribution, instead feed in the observations).

Further, the mutual information values calculated from Sklearns implementation are scaled by the natural logarithm of 2. In this implementation, the results are re-scaled to be linear.

Parameters

Optional[bool] = True

Save result to disk or not. Optional, default is to save.

Source code in key_interactions_finder/stat_modelling.py

def calc_mutual_info_to_target(self, save_result: bool = True):
    """
    Calculate the mutual information between each feature to the 2 target classes.
    Note that Sklearns implementation (used here) is designed for "raw datasets"
    (i.e., do not feed in a probability distribution, instead feed in the observations).

    Further, the mutual information values calculated from Sklearns implementation are
    scaled by the natural logarithm of 2. In this implementation,
    the results are re-scaled to be linear.

    Parameters
    ----------

    save_result : Optional[bool] = True
        Save result to disk or not.
        Optional, default is to save.
    """
    df_features = self.scaled_dataset.drop("Target", axis=1)
    features_array = df_features.to_numpy()
    classes = self.scaled_dataset["Target"].to_numpy()

    mutual_info_raw = mutual_info_classif(features_array, classes)
    mutual_info_rescaled = np.around((np.exp(mutual_info_raw) - 1), 5)

    self.mutual_infos = dict(
        zip(df_features.columns, mutual_info_rescaled))
    self.mutual_infos = {k: v for k, v in sorted(
        self.mutual_infos.items(), key=lambda item: item[1], reverse=True)}

    print("Mutual information scores calculated.")

    if save_result:
        out_file_path = Path(
            self.out_dir, "Mutual_Information_Per_Feature_Scores.csv")
        self._per_feature_scores_to_file(
            per_feat_values=self.mutual_infos,
            out_file=out_file_path
        )
        print(
            "You can also access these results via the class attribute: 'mutual_infos'.")

RegressionStatModel dataclass

Bases: _ProteinStatModel

Handles the generation of statistical models for PyContact data sets when the target variable is continous.

Note that several attributes listed below are inherited from _ProteinStatModel.

Attributes

pd.DataFrame

Input dataframe.

str

Directory path to store results files to. Default = ""

list, optional

What types of molecular interactions to generate the correlation matrix for. options are one or more of: ["Hbond", "Hydrophobic", "Saltbr", "Other"] Default is to include all 4 types.

pd.DataFrame

Input dataset with all features scaled.

list

List of all feature labels in the dataset.

dict

Dictionary with each feature's (keys) and mutual informations (values). Dictionary is sorted from largest mutual information to smallest.

dict

Dictionary with each feature's (keys) and linear correlations (values). Dictionary is sorted from largest (absolute) linear correlation to smallest.

Methods

calc_mutual_info_to_target(save_result=True) Calculate the mutual information between each feature and the target.

calc_linear_correl_to_target(save_result=True) Calculate the pearson correlation coeffcient between each feature and the target.

Source code in key_interactions_finder/stat_modelling.py

@dataclass
class RegressionStatModel(_ProteinStatModel):
    """
    Handles the generation of statistical models for PyContact data sets
    when the target variable is continous.

    Note that several attributes listed below are inherited from _ProteinStatModel.

    Attributes
    ----------

    dataset : pd.DataFrame
        Input dataframe.

    out_dir : str
        Directory path to store results files to.
        Default = ""

    interaction_types_included : list, optional
        What types of molecular interactions to generate the correlation matrix for.
        options are one or more of: ["Hbond", "Hydrophobic", "Saltbr", "Other"]
        Default is to include all 4 types.

    scaled_dataset : pd.DataFrame
        Input dataset with all features scaled.

    feature_list : list
        List of all feature labels in the dataset.

    mutual_infos : dict
        Dictionary with each feature's (keys) and mutual informations (values).
        Dictionary is sorted from largest mutual information to smallest.

    linear_correlations : dict
        Dictionary with each feature's (keys) and linear correlations (values).
        Dictionary is sorted from largest (absolute) linear correlation to smallest.

    Methods
    -------

    calc_mutual_info_to_target(save_result=True)
        Calculate the mutual information between each feature and the target.

    calc_linear_correl_to_target(save_result=True)
        Calculate the pearson correlation coeffcient between each feature and the target.
    """

    # Attribute is generated after initiziliation.
    linear_correlations: dict = field(init=False)

    # Called at the end of the dataclass's initialization procedure.
    def __post_init__(self) -> None:
        """Filter and rescale the features."""
        self.out_dir = _prep_out_dir(self.out_dir)

        self.mutual_infos = {}
        self.linear_correlations = {}

        if sorted(self.interaction_types_included) != sorted(
                ["Hbond", "Hydrophobic", "Saltbr", "Other"]):
            self.dataset = _filter_features_by_strings(
                dataset=self.dataset,
                strings_to_preserve=self.interaction_types_included
            )

        self.feature_list = list(self.dataset.columns)
        self.feature_list.remove("Target")

        # Features need to be scaled in order to use same bandwidth throughout.
        self.scaled_dataset = self._scale_features()

    def calc_mutual_info_to_target(self, save_result: bool = True) -> None:
        """
        Calculate the mutual information between each feature and the target.
        The target variable should be continuous.
        Note that Sklearns implementation (used here) is designed for "raw datasets"
        (i.e., do not feed in a probability distribution, instead feed in the observations).

        Further, the mutual information values calculated from Sklearns implementation are
        scaled by the natural logarithm of 2. In this implementation,
        the results are re-scaled to be linear.

        Parameters
        ----------

        save_result : Optional[bool] = True
            Save result to disk or not.
            Optional, default is to save.
        """
        df_features = self.scaled_dataset.drop("Target", axis=1)
        features_array = df_features.to_numpy()
        target_values = self.scaled_dataset["Target"].to_numpy()

        mutual_info_raw = mutual_info_regression(features_array, target_values)
        mutual_info_rescaled = np.around((np.exp(mutual_info_raw) - 1), 5)

        self.mutual_infos = dict(
            zip(df_features.columns, mutual_info_rescaled))
        self.mutual_infos = {k: v for k, v in sorted(
            self.mutual_infos.items(), key=lambda item: item[1], reverse=True)}

        print("Mutual information scores calculated.")

        if save_result:
            out_file_path = Path(
                self.out_dir, "Mutual_Information_Per_Feature_Scores.csv")
            self._per_feature_scores_to_file(
                per_feat_values=self.mutual_infos,
                out_file=out_file_path
            )

            print(
                "You can also access these results via the class attribute: 'mutual_infos'.")

    def calc_linear_correl_to_target(self, save_result: bool = True) -> None:
        """
        Calculate the pearson correlation coeffcient between each feature and the target.

        Parameters
        ----------

        save_result : Optional[bool] = True
            Save result to disk or not.
            Optional, default is to save.
        """
        target = self.dataset["Target"]
        features = self.dataset.drop(["Target"], axis=1)

        correlations = features.corrwith(target).to_frame()
        sorted_correlations = correlations.sort_values(
            by=0, key=abs, ascending=False)
        self.linear_correlations = sorted_correlations.to_dict(
            orient='dict')[0]

        print("Linear correlations calculated.")

        if save_result:
            out_file_path = Path(
                self.out_dir, "Linear_Correlations_Per_Feature_Scores.csv")
            self._per_feature_scores_to_file(
                per_feat_values=self.linear_correlations,
                out_file=out_file_path
            )

            print(
                "You can also access these results via the class attribute: 'linear_correlations'.")

__post_init__()

Filter and rescale the features.

Source code in key_interactions_finder/stat_modelling.py

def __post_init__(self) -> None:
    """Filter and rescale the features."""
    self.out_dir = _prep_out_dir(self.out_dir)

    self.mutual_infos = {}
    self.linear_correlations = {}

    if sorted(self.interaction_types_included) != sorted(
            ["Hbond", "Hydrophobic", "Saltbr", "Other"]):
        self.dataset = _filter_features_by_strings(
            dataset=self.dataset,
            strings_to_preserve=self.interaction_types_included
        )

    self.feature_list = list(self.dataset.columns)
    self.feature_list.remove("Target")

    # Features need to be scaled in order to use same bandwidth throughout.
    self.scaled_dataset = self._scale_features()

calc_linear_correl_to_target(save_result=True)

Calculate the pearson correlation coeffcient between each feature and the target.

Parameters

Optional[bool] = True

Save result to disk or not. Optional, default is to save.

Source code in key_interactions_finder/stat_modelling.py

def calc_linear_correl_to_target(self, save_result: bool = True) -> None:
    """
    Calculate the pearson correlation coeffcient between each feature and the target.

    Parameters
    ----------

    save_result : Optional[bool] = True
        Save result to disk or not.
        Optional, default is to save.
    """
    target = self.dataset["Target"]
    features = self.dataset.drop(["Target"], axis=1)

    correlations = features.corrwith(target).to_frame()
    sorted_correlations = correlations.sort_values(
        by=0, key=abs, ascending=False)
    self.linear_correlations = sorted_correlations.to_dict(
        orient='dict')[0]

    print("Linear correlations calculated.")

    if save_result:
        out_file_path = Path(
            self.out_dir, "Linear_Correlations_Per_Feature_Scores.csv")
        self._per_feature_scores_to_file(
            per_feat_values=self.linear_correlations,
            out_file=out_file_path
        )

        print(
            "You can also access these results via the class attribute: 'linear_correlations'.")

calc_mutual_info_to_target(save_result=True)

Calculate the mutual information between each feature and the target. The target variable should be continuous. Note that Sklearns implementation (used here) is designed for "raw datasets" (i.e., do not feed in a probability distribution, instead feed in the observations).

Further, the mutual information values calculated from Sklearns implementation are scaled by the natural logarithm of 2. In this implementation, the results are re-scaled to be linear.

Parameters

Optional[bool] = True

Save result to disk or not. Optional, default is to save.

Source code in key_interactions_finder/stat_modelling.py

def calc_mutual_info_to_target(self, save_result: bool = True) -> None:
    """
    Calculate the mutual information between each feature and the target.
    The target variable should be continuous.
    Note that Sklearns implementation (used here) is designed for "raw datasets"
    (i.e., do not feed in a probability distribution, instead feed in the observations).

    Further, the mutual information values calculated from Sklearns implementation are
    scaled by the natural logarithm of 2. In this implementation,
    the results are re-scaled to be linear.

    Parameters
    ----------

    save_result : Optional[bool] = True
        Save result to disk or not.
        Optional, default is to save.
    """
    df_features = self.scaled_dataset.drop("Target", axis=1)
    features_array = df_features.to_numpy()
    target_values = self.scaled_dataset["Target"].to_numpy()

    mutual_info_raw = mutual_info_regression(features_array, target_values)
    mutual_info_rescaled = np.around((np.exp(mutual_info_raw) - 1), 5)

    self.mutual_infos = dict(
        zip(df_features.columns, mutual_info_rescaled))
    self.mutual_infos = {k: v for k, v in sorted(
        self.mutual_infos.items(), key=lambda item: item[1], reverse=True)}

    print("Mutual information scores calculated.")

    if save_result:
        out_file_path = Path(
            self.out_dir, "Mutual_Information_Per_Feature_Scores.csv")
        self._per_feature_scores_to_file(
            per_feat_values=self.mutual_infos,
            out_file=out_file_path
        )

        print(
            "You can also access these results via the class attribute: 'mutual_infos'.")