ccle_data.py

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

'''
This code reproduces the real data experiments using the CCLE data.
Preprocessing steps follow the code of W. Tansey at https://github.com/tansey/hrt.
And the following code is obtained from https://github.com/alexisbellot/GCIT. 
'''


def load_ccle(drug_target='PLX4720', feature_type='both', normalize=False):
    '''
    :param drug target: specific drug we w ant to analyse
    :param normalize: normalize data
    :return: genetic features (mutations) as a 2d array for each cancer cell and corresponding drug response measured with Amax
    '''
    if feature_type in ['expression', 'both']:
        # Load gene expression
        expression = pd.read_csv('./ccle_data/expression.txt', delimiter='\t', header=2, index_col=1).iloc[:, 1:]
        expression.columns = [c.split(' (ACH')[0] for c in expression.columns]
        features = expression
    if feature_type in ['mutation', 'both']:
        # Load gene mutation
        mutations = pd.read_csv('./ccle_data/mutation.txt', delimiter='\t', header=2, index_col=1).iloc[:,1:]
        mutations = mutations.iloc[[c.endswith('_MUT') for c in mutations.index]]
        features = mutations
    if feature_type == 'both':
        # get cells having both expression and mutation data
        both_cells = set(expression.columns) & set(mutations.columns)
        z = {}
        for c in both_cells:
            exp = expression[c].values
            if len(exp.shape) > 1:
                exp = exp[:, 0]
            z[c] = np.concatenate([exp, mutations[c].values])
        both_df = pd.DataFrame(z, index=[c for c in expression.index] + [c for c in mutations.index])
        features = both_df

    print('Genetic features dimension = {} on {} cancer cells'.format(features.shape[0], features.shape[1]))

    # Get per-drug X and y regression targets
    response = pd.read_csv('./ccle_data/response.csv', header=0, index_col=[0, 2])

    # names of cell lines, there are 504
    cells = response.index.levels[0]
    # names of drugs, there are 24
    drugs = response.index.levels[1]

    X_drugs = [[] for _ in drugs]
    y_drugs = [[] for _ in drugs]

    # subset data to include only cells, mutations and response associated with chosen drug
    for j, drug in enumerate(drugs):
            if drug_target is not None and drug != drug_target:
                continue # return to beginning of the loop
            for i, cell in enumerate(cells):
                if cell not in features.columns or (cell, drug) not in response.index:
                    continue
                # all j empty except index that corresponds to target drug
                # for this j we iteratively append all the mutations on every cell
                X_drugs[j].append(features[cell].values) # store genetic features (mutations and expression) that appear in cells
                y_drugs[j].append(response.loc[(cell, drug), 'Amax']) # store response of the drug
            print('{}: Cell number = {}'.format(drug, len(y_drugs[j])))

    # convert to np array
    X_drugs = [np.array(x_i) for x_i in X_drugs]
    y_drugs = [np.array(y_i) for y_i in y_drugs]

    if normalize:
        X_drugs = [(x_i if (len(x_i) == 0) else (x_i - x_i.min(axis=0, keepdims=True)) /
                                                (x_i.max(axis=0, keepdims=True) - x_i.min(axis=0, keepdims=True)))
                   for x_i in X_drugs]
        y_drugs = [(y_i if (len(y_i) == 0 or y_i.std() == 0) else (y_i - y_i.min(axis=0, keepdims=True)) /
                                                (y_i.max(axis=0, keepdims=True) - y_i.min(axis=0, keepdims=True)))
                   for y_i in y_drugs]

    '''
    if normalize:
        X_drugs = [(x_i if (len(x_i) == 0) else (x_i - x_i.mean(axis=0, keepdims=True)) / 
        x_i.std(axis=0).clip(1e-6)) for x_i in X_drugs]
        y_drugs = [(y_i if (len(y_i) == 0 or y_i.std() == 0) else (y_i - y_i.mean()) / y_i.std()) for y_i in y_drugs]
    '''
    drug_idx = drugs.get_loc(drug_target)
    # 2d array for features and 1d array for response
    X_drug, y_drug = X_drugs[drug_idx], y_drugs[drug_idx]

    return X_drug, y_drug, features


# X_drug, y_drug, features = load_ccle(feature_type='mutation')


def ccle_feature_filter(X, y, threshold=0.1):
    '''
    :param X: features
    :param y: response
    :param threshold: correlation threshold
    :return: logical array with False for all features that do not have at least pearson correlation at threshold with y
    and correlations for all variables
    '''
    corrs = np.array([np.abs(np.corrcoef(x, y)[0, 1]) if x.std() > 0 else 0 for x in X.T])
    selected = corrs >= threshold # True/False
    print(selected.sum(), selected.shape, corrs)
    return selected, corrs

# ccle_selected, corrs = ccle_feature_filter(X_drug, y_drug, threshold=0.1)

# features.index[ccle_selected]
# stats.describe(corrs[ccle_selected])


def fit_elastic_net_ccle(X, y, nfolds=3):
    '''
    :param X: features
    :param y: response
    :param nfolds: number of folds for hyperparameter tuning
    :return: fitted elastic net model
    '''
    from sklearn.linear_model import ElasticNetCV
    # The parameter l1_ratio corresponds to alpha in the glmnet R package
    # while alpha corresponds to the lambda parameter in glmnet
    # enet = ElasticNetCV(l1_ratio=np.linspace(0.2, 1.0, 10),
    #                     alphas=np.exp(np.linspace(-6, 5, 250)),
    #                     cv=nfolds)
    enet = ElasticNetCV(l1_ratio=0.2, # It always chooses l1_ratio=0.2
                        alphas=np.exp(np.linspace(-6, 5, 250)),
                        cv=nfolds)
    print('Fitting via CV')
    enet.fit(X, y)
    alpha, l1_ratio = enet.alpha_, enet.l1_ratio_
    print('Chose values: alpha={}, l1_ratio={}'.format(alpha, l1_ratio))
    return enet

# elastic_model = fit_elastic_net_ccle(X_drug[:,ccle_selected], y_drug)


def fit_random_forest_ccle(X, y):
    '''
    :param X: features
    :param y: response
    :param nfolds: number of folds for hyperparameter tuning
    :return: fitted elastic net model
    '''
    from sklearn.ensemble import RandomForestRegressor

    rf = RandomForestRegressor()

    rf.fit(X,y)

    return rf

# rf_model = fit_random_forest_ccle(X_drug[:,ccle_selected], y_drug)


def plot_ccle_predictions(model, X, y):
    from sklearn.metrics import r2_score
    plt.close()
    y_hat = model.predict(X)
    plt.scatter(y_hat, y, color='blue')
    plt.plot([min(y.min(), y_hat.min()), max(y.max(), y_hat.max())],
             [min(y.min(), y_hat.min()),max(y.max(), y_hat.max())], color='red', lw=3)
    plt.xlabel('Predicted')
    plt.ylabel('Truth')
    plt.title(' ($r^2$={:.4f})'.format( r2_score(y, y_hat)))
    plt.tight_layout()

# plot_ccle_predictions(elastic_model,X_drug[:,ccle_selected],y_drug)


def print_top_features(model):
    # model_weights = np.mean([m.coef_ for m in model.models], axis=0)
    if model == rf_model:
        model_weights = model.feature_importances_
    else:
        model_weights = model.coef_

    ccle_features = features[ccle_selected]

    print('Top by fit:')
    for idx, top in enumerate(np.argsort(np.abs(model_weights))[::-1]):
        print('{}. {}: {:.4f}'.format(idx+1, ccle_features.index[top], model_weights[top]))

# print_top_features(rf_model)
# print_top_features(elastic_model)


def run_test_ccle(X, Y):
    pval = []
    for x_index in range(X.shape[1]):
        z = np.delete(X, x_index, axis=1)
        x = X[:, x_index]
        x = x.reshape((len(x), 1))
        Y = Y.reshape((len(Y), 1))
        # now run test
        pval.append(GCIT(x, Y, z))

    ccle_features = features[ccle_selected]

    print('Top by fit:')
    for idx, top in enumerate(np.argsort(np.abs(pval))):
        print('{}. {}: {:.4f}'.format(idx+1, ccle_features.index[top], pval[top]))

# run_test_ccle(X_drug[:,ccle_selected],y_drug)