SKM_cluster.py

import os, sys
import argparse
import pandas as pd
import pandas.rpy.common as com
import rpy2
import random
import subprocess


def r_installed():
    """ check that r is installed on system
    does not work on Windows """
    try:
        subprocess.check_output(['which','R'])
        return True
    except:
        if 'win' in sys.platform:
            print 'Warning: unable to check for R in windows'
        return False

 
def run_command(cmd):
    """ run command, return stdout, stderr, returncode"""
    try:
        retcode = subprocess.call(cmd, shell=True)
        if retcode < 0:
            print >>sys.stderr, "Child was terminated by signal", -retcode
        else:
            print >>sys.stderr, "Child returned", retcode
        return retcode
    except OSError as e:
        print >>sys.stderr, "Execution failed:", e
        return None

def get_sparcl_dir():
    """ returns default location of sparcl library dir for this user"""
    home = os.environ['HOME']
    return os.path.join(home, '.rpackages', 'sparcl')
    

def sparcl_installed():
    """ check if sparcl library is installed in default location
    (default: /userhome/.rpackages/sparcl) """   
    return os.path.isdir(get_sparcl_dir())

def install_sparcl():
    home = os.environ['HOME']
    rpkg_dir = os.path.join(home, '.rpackages')
    if not os.path.isdir(rpkg_dir):
        os.mkdir(rpkg_dir)
    curr_dir, _ = os.path.split(__file__)
    sparcl_loc = os.path.join(curr_dir, 'sparcl')
    cmd = ' '.join(['R', 'CMD', 'INSTALL', sparcl_loc, '-l', rpkg_dir])
    run_command(cmd)
    

def import_sparcl():
    if not sparcl_installed():
        install_sparcl()
    sparcl_dir = get_sparcl_dir()
    rpkg_dir, _ = os.path.split(sparcl_dir)
    sparcl = rpy2.robjects.packages.importr("sparcl", lib_loc=rpkg_dir)
    return sparcl


def create_rslts_frame(dataframe):  
    """ Create empty pandas dataframe to hold feature weights and cluster 
    membership results of each resampling run""" 
    weight_rslts = pd.DataFrame(data=None, index = dataframe.columns)
    clust_rslts = pd.DataFrame(data=None, index = dataframe.index)
    return weight_rslts, clust_rslts
    
    
def sample_data(data, split = .7):
    """
    Takes an array of data as input. Randomly samples 70% of the observations
    and returns as an array
    """
    samp_n = int(split * len(data))
    rand_samp = sorted(random.sample(xrange(len(data)), samp_n))
    sampdata = data.take(rand_samp)
    unsamp_idx = [x for x in xrange(len(data)) if x not in rand_samp]
    unsampdata = data.take(unsamp_idx)
    return sampdata, unsampdata


def skm_permute(data):
    """
    rpy2 wrapper for R function: KMeansSparseCluster.permute from the sparcl package.
    The tuning parameter controls the L1 bound on w, the feature weights. A permutation 
    approach is used to select the tuning parameter. 
    
    Infile:
    ---------------
    data: pandas Dataframe
            nxp dataframe where n is observations and p is features (i.e. ROIs)
            Should be a pandas DataFrame with subject codes as index and features 
            as columns.
    
    Returns:
    ---------------
    best_L1bound: float
    	tuning parameter that returns the highest gap statistic
	(more features given non-zero weights)
    
    lowest_L1bound: float
    	smallest tuning parameter that gives a gap statistic within 
        one sdgap of the largest gap statistic (sparser result)
    """
    sparcl = import_sparcl()
    r_data = com.convert_to_r_dataframe(data)
    km_perm = sparcl.KMeansSparseCluster_permute(r_data,K=2,nperms=25)
    best_L1bound = km_perm.rx2('bestw')[0]
    wbounds = km_perm.rx2('wbounds')
    gaps = km_perm.rx2('gaps')
    bestgap = max(gaps)
    sdgaps = km_perm.rx2('sdgaps')
    # Calculate smallest wbound that returns gap stat within one sdgap of best wbound
    wbound_rnge = [wbounds[i] for i in range(len(gaps)) if (gaps[i]+sdgaps[i]>=bestgap)]
    lowest_L1bound = min(wbound_rnge)
    return best_L1bound, lowest_L1bound
    
    
def skm_cluster(data, L1bound):
    """
    rpy2 wrapper for R function: KMeansSparseCluster from the sparcl package.
    This function performs sparse k-means clustering. You must specify L1 bound on w, 
    the feature weights. 
    
    Note: A smaller L1 bound will results in sparser weighting. If a large number of 
    features are included, it may be useful to use the smaller tuning parameter 
    returned by KMeansSparseCluster.permute wrapper function.
    
    Infile:
    ---------------
    data: pandas Dataframe
            nxp dataframe where n is observations and p is features (i.e. ROIs)
            Should be a pandas DataFrame with subject codes as index and features 
            as columns.
    
    Returns:
    ---------------
    km_weights: pandas DataFrame   
                index = feature labels, values = weights
    km_clusters: pandas DataFrame
                index = subject code, values = cluster membership
    """
    sparcl = import_sparcl()
    # Convert pandas dataframe to R dataframe
    r_data = com.convert_to_r_dataframe(data)
    # Cluster observations using specified L1 bound
    km_out = sparcl.KMeansSparseCluster(r_data,K=2, wbounds=L1bound)
    # Create dictionary of feature weights, normalized feature weights and
    # cluster membership
    ws = km_out.rx2(1).rx2('ws')
    km_weights = {k.replace('.','-'): [ws.rx2(k)[0]] for k in ws.names}
    km_weights = pd.DataFrame.from_dict(km_weights)
    km_weightsT = km_weights.T
    km_weightsnorm = km_weightsT/km_weightsT.sum()
    Cs = km_out.rx2(1).rx2('Cs')
    km_clusters = {k: [Cs.rx2(k)[0]] for k in Cs.names}   
    km_clusters = pd.DataFrame.from_dict(km_clusters)
    km_clusters = km_clusters.T
    return km_weightsnorm, km_clusters
        
        
def calc_cutoffs(data, weights, weightsum, clusters):
    """
    Determines the cluster centers for regions. This will only detmine cutoffs 
    for features with weights making up the top percentile specified. Clusters 
    are then labelled as positive or negative based on which has the highest 
    average PIB index across these features. This is done because clustering 
    does not consistently assign the same label to groups across runs. 
    
    Inputs
    ----------
    data:   pandas Dataframe
            nxp dataframe where n is observations and p is features (i.e. ROIs)
            Should be a pandas DataFrame with subject codes as index and features 
            as columns.
            
    weights:    pandas Dataframe   
                index = feature labels, values = weights
    weightsum:  float
                cluster cut-offs will be determined for only for features with weihts
                in the top percentile specified
    clusters:   pandas Dataframe
                index = subject code, values = cluster membership
    
    Returns:
    ----------
    clust_renamed:  pandas Dataframe
                    renamed version of clusters input such that integer labels
                    are replaced with appropriate string label ('pos' or 'neg')
    
    cutoffs:    pandas Dataframe
                index = features labels accounting for to 50% of weights
                values = cut-off score determined as mid-point of cluster means
                        for each feature
    
    Note:
    The cluster means may be used to create cut-off scores (i.e. the midpoint 
    between cluster means) in order to predict cluster membership of other 
    subjects. If a large number of features were used, it may be useful to 
    set a lower weightsum (i.e., .50). This will constrain the features used
    to classify subjects to only those that contributed most to clustering. 
    """
    # Select features accounting for up to 50% of the weighting
    sorted_weights = weights.sort(columns=0, ascending=False)
    sum_weights = sorted_weights.cumsum()
    topfeats = sum_weights[sum_weights[0] <= weightsum].index
    clust1subs = clusters[clusters[0] == 1].index
    clust2subs = clusters[clusters[0] == 2].index
    clust1dat = data.reindex(index=clust1subs, columns=topfeats)
    clust2dat = data.reindex(index=clust2subs, columns=topfeats)
    clust1mean = clust1dat.mean(axis=0)
    clust2mean = clust2dat.mean(axis=0)
    if clust1mean.mean() > clust2mean.mean():
        pos_means = clust1mean
        neg_means = clust2mean
        clust_renamed = clusters[0].astype(str).replace(['1','2'], ['pos','neg'])
    elif clust1mean.mean() < clust2mean.mean():
        pos_means = clust2mean
        neg_means = clust1mean
        clust_renamed = clusters[0].astype(str).replace(['1','2'], ['neg','pos'])
    cutoffs = (pos_means + neg_means) / 2
    clust_renamed.name = 'PIB_Status'
    return clust_renamed, cutoffs


def predict_clust(data, cutoffs):
    """
    Predict cluster membership for set of subjects using feature cutoff scores.
    Classifies as postive if any feature surpasses cut-off value.
    
    Inputs:
    -------------
    data:   pandas DataFrame  
            nxp dataframe where n is observations and p is features (i.e. ROIs)
            Should include subject codes as index and features 
            as columns.
            
    cutoffs:   pandas DataFrame   
                index = features of interest labels, values = weights 
    
    Returns:
    predicted_clust:    pandas DataFrame
                        index = subject codes passed from input data
                        values = predicted cluster membership     
    """
    # Check all features against cut-offs
    cutdata = data[cutoffs.index] > cutoffs 
    # Classify as pos if any feature is above cutoff
    cutdata_agg = cutdata.any(axis=1)
    predicted_clust = cutdata_agg.astype(str).replace(['T', 'F'], ['pos', 'neg'])
    predicted_clust.name = 'PIB_Status'
    return predicted_clust
    
    
def create_tighclust(clusterdata):
    """
    Creates tight clusters consisting of subjects classified as a member of given 
    cluster in at least 96% of resample runs.
    
    Inputs:
    ------------
    clusterdata: pandas Dataframe
                    Dataframe containing cluster membership over all resamples
                    index = subject code 
                    columns = resample run 
                    values = cluster membership
    Returns:
    ------------
    tight_subs: pandas Dataframe 
                Dataframe containing only subjects belonging to tight clusters
                and their cluster membership
    """
        
    clust_totals = clusterdata.apply(pd.value_counts, axis=1)
    clust_pct = clust_totals / len(clusterdata.columns)
    pos_subs = clust_pct.index[clust_pct['pos'] > .96]
    neg_subs = clust_pct.index[clust_pct['neg'] > .96]
    tight_subs = pd.DataFrame(index=pos_subs + neg_subs,columns=pd.Index([0]))
    tight_subs[0].ix[pos_subs] = 1
    tight_subs[0].ix[neg_subs] = 2
    return tight_subs
    

def run_clustering(infile, nperm, weightsum, bound):
    """
    Runs sparse k-means resampling. For each sample, runs a clustering on a
    subset of subjects. These clusters of subjects are used to generate 
    regional cut-offs in order to classify the remaining subjects in the sample.
    
    Inputs:
    ---------
    infile: str
            path to input datafile. First column should contain subject
            codes and additional columns should correspond to features
    nperm: int
            number of clustering resample runs
    weightsum: float
                percentage of total feature weights to use in calculating cutoffs
                Only features with the highest weights summing to this value will 
                be used.
    bound: str  ['best' or 'sparse']
            Determines which value generated by SKM permutation to use as
            L1 bound in clustering. This tuning parameter determines how weights 
            will be distributed among feratures. 'best' will give more non-zero 
            weights.   
    
    Returns:
    ----------
    weight_rslts: pandas Dataframe
                    nxp dataframe where n is the number of features and p is
                    the number of nperms. 
                    Index = feature names, values = feature weights
    clust_rslts: pandas Dataframe
                    nxp dataframe where n is the number of subject and p is
                    the number of nperms. 
                    Index = subject codes, values = cluster membership
            
    """
    # make sure we can import sparcl
    import_sparcl()
    # Load data to dataframe
    dataframe = pd.read_csv(infile, sep=None, index_col=0)
    # Create empty frames to hold results of resampling
    weight_rslts, clust_rslts = create_rslts_frame(dataframe) 
    for resamp_run in range(nperm):
        print 'Now starting re-sample run number %s'%(resamp_run)
        # Get random sub-sample (without replacement) of group to feed into clustering
        # Currently set to 70% of group N
        traindat, testdat = sample_data(dataframe)
        best_L1bound, lowest_L1bound = skm_permute(traindat)
        if bound == 'sparse':
            km_weight, km_clust = skm_cluster(traindat, lowest_L1bound)      
        else:
            km_weight, km_clust = skm_cluster(traindat, best_L1bound)
        samp_clust, sampcutoffs = calc_cutoffs(traindat, km_weight, weightsum, km_clust)
        unsamp_clust = predict_clust(testdat, sampcutoffs)
        # Log weights and cluster membership of resample run
        weight_rslts[resamp_run] = km_weight[0]
        clust_rslts[resamp_run] = pd.concat([samp_clust, unsamp_clust])
    return weight_rslts, clust_rslts
    

def classify_subjects(infile, weightsum, weight_rslts, clust_rslts):
    """
    Classifies subject into PIB+ and PIB- based on regional cut-offs
    derived from sparse k-means clustering. 
    
    Inputs:
    --------
    infile: str
            path to input datafile. First column should contain subject
            codes and additional columns should correspond to features
    weightsum: float
                percentage of total feature weights to use in calculating cutoffs
                Only features with the highest weights summing to this value will 
                be used.
    weight_rslts: pandas Dataframe
                    nxp dataframe where n is the number of features and p is
                    the number of nperms. 
                    Index = feature names, values = feature weights
    clust_rslts: pandas Dataframe
                    nxp dataframe where n is the number of subject and p is
                    the number of nperms. 
                    Index = subject codes, values = cluster membership
                    
    Returns:
    ---------
    all_clust: pandas Dataframe
                    Contains cluster membership for all subjects
                    index = subject codes passed from input data
                    values = predicted cluster membership  
    weight_totals: pandas Dataframe
                        Contains average weight for all feature s
                        index = feature names
                        values = average weight
    grpcutoffs: pandas Dataframe
                    Contains PIB value used as cutoff for all features
                    index = feature names
                    values = value used as cut-off between groups
    """
    # make sure we can import sparcl
    import_sparcl()
    # Load data to dataframe
    dataframe = pd.read_csv(infile, sep=None, index_col=0)
    # Create tight clusters and generate regional cut-offs to predict remaining subjects
    weight_totals = weight_rslts.mean(axis=1)
    tight_subs = create_tighclust(clust_rslts)
    tight_clust, grpcutoffs = calc_cutoffs(dataframe.ix[tight_subs.index], 
                                            pd.DataFrame(weight_totals), 
                                            weightsum,
                                            tight_subs)
    all_clust = predict_clust(dataframe, grpcutoffs) 
    grpcutoffs.name = 'Cutoff_Value'
    weight_totals.name = 'WeightMean'    
    return all_clust, grpcutoffs, weight_totals
    
def save_results(outdir, all_clust, weight_totals, grpcutoffs):
    """
    Inputs:
    ---------
    outdir: str
                Path to save output files
    all_clust: pandas Dataframe
                    Contains cluster membership for all subjects
                    index = subject codes passed from input data
                    values = predicted cluster membership  
    weight_totals: pandas Dataframe
                        Contains average weight for all feature s
                        index = feature names
                        values = average weight
    grpcutoffs: pandas Dataframe
                    Contains PIB value used as cutoff for all features
                    index = feature names
                    values = value used as cut-off between groups    
    
    Returns:
    ---------
    all_clust_out: str
                    path to ClusterResults.csv, contains cluster membership 
                    of all subjects
    weight_totals_out: str
                        path to FeatureWeights.csv. contains average weight
                        for each feature
    grpcutoffs_out: str
                        path to CutoffValues.csv, contains cutoff value for
                        each feature used to classify subjects
    """
    
    # Save out results of cluster membership and feature weights
    all_clust_out = os.path.join(outdir, 'ClusterResults.csv')
    all_clust.to_csv(all_clust_out, index=True, 
                        index_label='SUBID', header=True,sep='\t')
    print 'Cluster membership results saved to %s'%(all_clust_out)
    
    weight_totals_out = os.path.join(outdir, 'FeatureWeights.csv')
    weight_totals.to_csv(weight_totals_out, index=True, 
                            index_label='Feature', header=True,sep='\t')
    print 'Mean feature weights saved to %s'%(weight_totals_out)
    
    grpcutoffs_out = os.path.join(outdir, 'CutoffValues.csv')
    grpcutoffs.to_csv(grpcutoffs_out, index=True, 
                            index_label='Feature', header=True,sep='\t')
    print 'Regions and cutoff scores used to determine groups saved to %s'%(grpcutoffs_out)
    return all_clust_out, weight_totals_out, grpcutoffs_out
    
    
def main(infile, outdir, nperm, weightsum, bound):
    """
    Function to run full script. Takes input from command line and runs 
    sparse k-means clustering with resampling to produce tight clusters.
    Regional cutoffs are derived from these tight clusters and subjects
    are classified based on these cutoffs. Saves out results to file.
    """
    weight_rslts, clust_rslts = run_clustering(infile, nperm, weightsum, bound)
    all_clust, grpcutoffs, weight_totals = classify_subjects(infile, 
                                                            weightsum, 
                                                            weight_rslts, 
                                                            clust_rslts)
    all_clust_out, weight_totals_out, grpcutoffs_out = save_results(outdir, 
                                                                    all_clust, 
                                                                    weight_totals, 
                                                                    grpcutoffs)

##########################################################################
        
if __name__ == '__main__':

    parser = argparse.ArgumentParser(description = """
    Sparse K-means Clustering with re-sampling to cluster subjects into 
    PIB+ and PIB- groups.
    -------------------------------------------------------------------""",
                                    formatter_class=argparse.RawTextHelpFormatter,
                                    epilog= """
    -------------------------------------------------------------------
    NOTE: Format of infile should be a spreadsheet with first column 
    containing subject codes and the remaining columns corresponding to 
    ROIs to be used as features in clustering. The first row may contain 
    headers.""")

    parser.add_argument('infile', type=str, nargs=1,
            help='Input file containing subject codes and PIB indices')
    parser.add_argument('-outdir', dest='outdir', default=None,
            help='Directory to save results. (default = infile directory)')
    parser.add_argument('-nperm', type=int, dest = 'nperm', default = 1000, 
            help = 'Number of re-sample permutations (default 1000)')
    parser.add_argument('-weightsum', type=float, dest = 'weightsum', 
                        default = 1.0, 
            help = """Only determine cutoffs for features with weights making 
up the top percentile specified. Constrains the number 
of features used for classification to those that were 
most important in clustering (default = 1.0)""")
    parser.add_argument('-bound', type=str, choices=['best', 'sparse'],
                        dest='bound', default='best',
            help="""method to determine L1bound, best result, or best  
    sparse result""")
	

    if len(sys.argv) == 1:
        parser.print_help()
    else:
        args = parser.parse_args()
        if args.outdir is None:
            args.outdir, _ = os.path.split(args.infile[0])

            
        ### Begin running SKM clustering and resampling  
        main(args.infile[0], args.outdir, args.nperm, args.weightsum, args.bound)