Add cancer types to single-cancer mutation prediction models

04_add_cancer_types/run_add_cancer_classification.py

@@ -0,0 +1,257 @@
+"""
+Script to run "add cancer" experiments. The general idea is to add cancers
+in a particular order (either random or determined by cancer type similarity).
+
+The prediction goal is the same as in 02_cancer_type_classification
+experiments; that is, predicting presence/absence of a given mutation in
+a particular cancer type.
+
+"""
+import sys
+import argparse
+import itertools as it
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+import pancancer_evaluation.config as cfg
+from pancancer_evaluation.data_models.tcga_data_model import TCGADataModel
+from pancancer_evaluation.exceptions import (
+    NoTrainSamplesError,
+    NoTestSamplesError,
+    OneClassError,
+    ResultsFileExistsError
+)
+from pancancer_evaluation.utilities.classify_utilities import run_cv_cancer_type
+import pancancer_evaluation.utilities.data_utilities as du
+import pancancer_evaluation.utilities.file_utilities as fu
+
+def process_args():
+    p = argparse.ArgumentParser()
+    p.add_argument('--custom_genes', nargs='*', default=None,
+                   help='currently this needs to be a subset of top_50')
+    p.add_argument('--debug', action='store_true',
+                   help='use subset of data for fast debugging')
+    p.add_argument('--gene_set', type=str,
+                   choices=['top_50', 'vogelstein', 'custom'],
+                   default='top_50',
+                   help='choose which gene set to use. top_50 and vogelstein are '
+                        'predefined gene sets (see data_utilities), and custom allows '
+                        'any gene or set of genes in TCGA, specified in --custom_genes')
+    p.add_argument('--holdout_cancer_types', nargs='*', default=None,
+                   help='provide a list of cancer types to hold out, uses all '
+                        'cancer types in TCGA if none are provided')
+    p.add_argument('--how_to_add', type=str,
+                   choices=['random', 'similarity'],
+                   default='random',
+                   help='Method for choosing cancer types to add to the '
+                        'training dataset; see data model for details')
+    p.add_argument('--log_file', default=None,
+                   help='name of file to log skipped cancer types to')
+    p.add_argument('--num_folds', type=int, default=4,
+                   help='number of folds of cross-validation to run')
+    p.add_argument('--results_dir', default=cfg.results_dir,
+                   help='where to write results to')
+    p.add_argument('--seed', type=int, default=cfg.default_seed)
+    p.add_argument('--subset_mad_genes', type=int, default=cfg.num_features_raw,
+                   help='if included, subset gene features to this number of '
+                        'features having highest mean absolute deviation')
+    p.add_argument('--verbose', action='store_true')
+    args = p.parse_args()
+
+    if args.gene_set == 'custom':
+        if args.custom_genes is None:
+            p.error('must include --custom_genes when --gene_set=\'custom\'')
+        args.gene_set = args.custom_genes
+        del args.custom_genes
+    elif (args.gene_set != 'custom' and args.custom_genes is not None):
+        p.error('must use option --gene_set=\'custom\' if custom genes are included')
+
+    sample_info_df = du.load_sample_info(args.verbose)
+    tcga_cancer_types = list(np.unique(sample_info_df.cancer_type))
+    if args.holdout_cancer_types is None:
+        args.holdout_cancer_types = tcga_cancer_types
+    else:
+        not_in_tcga = set(args.holdout_cancer_types) - set(tcga_cancer_types)
+        if len(not_in_tcga) > 0:
+            p.error('some cancer types not present in TCGA: {}'.format(
+                ' '.join(not_in_tcga)))
+
+    args.results_dir = Path(args.results_dir).resolve()
+
+    if args.log_file is None:
+        args.log_file = Path(args.results_dir, 'log_skipped.tsv').resolve()
+
+    return args, sample_info_df
+
+
+if __name__ == '__main__':
+
+    # process command line arguments
+    args, sample_info_df = process_args()
+
+    # create results dir if it doesn't exist
+    args.results_dir.mkdir(parents=True, exist_ok=True)
+
+    # create empty log file if it doesn't exist
+    log_columns = [
+        'gene',
+        'cancer_type',
+        'shuffle_labels',
+        'skip_reason'
+    ]
+    if args.log_file.exists() and args.log_file.is_file():
+        log_df = pd.read_csv(args.log_file, sep='\t')
+    else:
+        log_df = pd.DataFrame(columns=log_columns)
+        log_df.to_csv(args.log_file, sep='\t')
+
+    tcga_data = TCGADataModel(sample_info=sample_info_df,
+                              seed=args.seed,
+                              subset_mad_genes=args.subset_mad_genes,
+                              verbose=args.verbose,
+                              debug=args.debug)
+
+    genes_df = tcga_data.load_gene_set(args.gene_set)
+
+    # we want to run mutation prediction experiments:
+    # - for signal and shuffled labels
+    #   (shuffled labels acts as our lower baseline)
+    # - for all genes in the given gene set
+    # - for all cancer types in the given holdout cancer types (or all of TCGA)
+    # - for all numbers of cancers to add to training set
+    #   (cfg.num_train_cancer_types)
+    for shuffle_labels in (False, True):
+
+        print('shuffle_labels: {}'.format(shuffle_labels))
+
+        progress_1 = tqdm(genes_df.iterrows(),
+                          total=genes_df.shape[0],
+                          ncols=100,
+                          file=sys.stdout)
+
+        for gene_idx, gene_series in progress_1:
+            gene = gene_series.gene
+            classification = gene_series.classification
+            progress_1.set_description('gene: {}'.format(gene))
+
+            gene_dir = fu.make_gene_dir(args.results_dir, gene, add_cancer=True)
+
+            progress_2 = tqdm(args.holdout_cancer_types,
+                              ncols=100,
+                              file=sys.stdout)
+
+            for test_cancer_type in progress_2:
+
+                progress_2.set_description('cancer type: {}'.format(test_cancer_type))
+                cancer_type_log_df = None
+
+                progress_3 = tqdm(cfg.num_train_cancer_types,
+                                  ncols=100,
+                                  file=sys.stdout)
+
+                for num_train_cancer_types in progress_3:
+
+                    progress_3.set_description('num train cancers: {}'.format(
+                        num_train_cancer_types))
+
+                    try:
+                        tcga_data.process_data_for_gene_and_cancer(gene,
+                                                                   classification,
+                                                                   test_cancer_type,
+                                                                   gene_dir,
+                                                                   num_train_cancer_types,
+                                                                   how_to_add=args.how_to_add,
+                                                                   shuffle_labels=shuffle_labels)
+                    except NoTrainSamplesError:
+                        if args.verbose:
+                            print('Skipping due to no train samples: gene {}, '
+                                  'cancer type {}'.format(gene, test_cancer_type),
+                                  file=sys.stderr)
+                        cancer_type_log_df = fu.generate_log_df(
+                            log_columns,
+                            [gene, test_cancer_type, shuffle_labels, 'no_train_samples']
+                        )
+                        continue
+
+                    try:
+                        # check if results file already exists, if not skip it
+                        check_file = fu.check_add_cancer_file(gene_dir,
+                                                              gene,
+                                                              test_cancer_type,
+                                                              num_train_cancer_types,
+                                                              args.how_to_add,
+                                                              args.seed,
+                                                              shuffle_labels)
+                    except ResultsFileExistsError:
+                        if args.verbose:
+                            print('Skipping because results file exists already: '
+                                  'gene {}, cancer type {}'.format(gene, test_cancer_type),
+                                  file=sys.stderr)
+                        cancer_type_log_df = fu.generate_log_df(
+                            log_columns,
+                            [gene, test_cancer_type, shuffle_labels, 'file_exists']
+                        )
+                        continue
+
+                    try:
+                        # run cross-validation for the given cancer type
+                        #
+                        # since we already filtered the dataset to the cancer
+                        # types of interest, we can just use this function with
+                        # the "pancancer" option (you can think of it as a a
+                        # pancancer model where the "universe" of all cancers
+                        # is limited by our previous filtering, kinda).
+                        results = run_cv_cancer_type(tcga_data,
+                                                     gene,
+                                                     test_cancer_type,
+                                                     sample_info_df,
+                                                     args.num_folds,
+                                                     use_pancancer=True,
+                                                     use_pancancer_only=False,
+                                                     shuffle_labels=shuffle_labels)
+                    except NoTrainSamplesError:
+                        if args.verbose:
+                            print('Skipping due to no train samples: gene {}, '
+                                  'cancer type {}'.format(gene, test_cancer_type),
+                                  file=sys.stderr)
+                        cancer_type_log_df = fu.generate_log_df(
+                            log_columns,
+                            [gene, test_cancer_type, shuffle_labels, 'no_train_samples']
+                        )
+                    except NoTestSamplesError:
+                        if args.verbose:
+                            print('Skipping due to no test samples: gene {}, '
+                                  'cancer type {}'.format(gene, test_cancer_type),
+                                  file=sys.stderr)
+                        cancer_type_log_df = fu.generate_log_df(
+                            log_columns,
+                            [gene, test_cancer_type, shuffle_labels, 'no_test_samples']
+                        )
+                    except OneClassError:
+                        if args.verbose:
+                            print('Skipping due to one holdout class: gene {}, '
+                                  'cancer type {}'.format(gene, test_cancer_type),
+                                  file=sys.stderr)
+                        cancer_type_log_df = fu.generate_log_df(
+                            log_columns,
+                            [gene, test_cancer_type, shuffle_labels, 'one_class']
+                        )
+                    else:
+                        # only save results if no exceptions
+                        fu.save_results_add_cancer(gene_dir,
+                                                   check_file,
+                                                   results,
+                                                   gene,
+                                                   test_cancer_type,
+                                                   tcga_data.y_df.DISEASE.unique(),
+                                                   num_train_cancer_types,
+                                                   args.how_to_add,
+                                                   args.seed,
+                                                   shuffle_labels)
+
+                    if cancer_type_log_df is not None:
+                        fu.write_log_file(cancer_type_log_df, args.log_file)
+

pancancer_evaluation/config.py

@@ -58,3 +58,11 @@
 cross_cancer_types = [
     'THCA', 'COAD', 'GBM', 'LGG', 'SKCM'
 ]
+
+# parameters for "add cancer" experiments
+
+# how many cancer types to add to target cancer
+# 0 = just use target cancer, -1 = use all cancers (pan-cancer model)
+num_train_cancer_types = [0, 1, 2, 4, -1]
+# similarity matrix to use for 'similarity' addition option
+similarity_matrix_file = data_dir / 'expression_confusion_matrix.tsv'