def run(args):
Utilities.maybe_create_folder(args.intermediate_folder)
Utilities.ensure_requisite_folders(args.output_prefix)
logging.info("Opening data")
p_ = re.compile(args.data_name_pattern)
f = [x for x in sorted(os.listdir(args.data_folder)) if p_.search(x)]
tissue_names = [p_.search(x).group(1) for x in f]
data = []
for i in range(0, len(tissue_names)):
logging.info("Loading %s", tissue_names[i])
data.append((tissue_names[i],
pq.ParquetFile(os.path.join(args.data_folder, f[i]))))
data = collections.OrderedDict(data)
available_data = {
x
for p in data.values() for x in p.metadata.schema.names
}
logging.info("Preparing output")
WEIGHTS_FIELDS = [
"gene", "rsid", "varID", "ref_allele", "eff_allele", "weight"
]
SUMMARY_FIELDS = [
"gene", "genename", "gene_type", "alpha", "n_snps_in_window",
"n.snps.in.model", "rho_avg", "pred.perf.R2", "pred.perf.pval"
]
Utilities.ensure_requisite_folders(args.output_prefix)
if args.skip_regression:
weights, summaries, covariances = None, None, None
else:
weights, summaries, covariances = setup_output(args.output_prefix,
tissue_names,
WEIGHTS_FIELDS,
SUMMARY_FIELDS)
logging.info("Loading data annotation")
data_annotation = StudyUtilities._load_gene_annotation(
args.data_annotation)
data_annotation = data_annotation[data_annotation.gene_id.isin(
available_data)]
if args.chromosome or (args.sub_batches and args.sub_batch):
data_annotation = StudyUtilities._filter_gene_annotation(
data_annotation, args.chromosome, args.sub_batches, args.sub_batch)
logging.info("Kept %i entries", data_annotation.shape[0])
logging.info("Opening features annotation")
if not args.chromosome:
features_metadata = pq.read_table(args.features_annotation).to_pandas()
else:
features_metadata = pq.ParquetFile(
args.features_annotation).read_row_group(args.chromosome -
1).to_pandas()
if args.discard_palindromic_snps:
logging.info("Discarding palindromic snps")
features_metadata = Genomics.discard_gtex_palindromic_variants(
features_metadata)
if args.chromosome and args.sub_batches:
logging.info("Trimming variants")
features_metadata = StudyUtilities.trim_variant_metadata_on_gene_annotation(
features_metadata, data_annotation, args.window)
if args.rsid_whitelist:
logging.info("Filtering features annotation")
whitelist = TextFileTools.load_list(args.rsid_whitelist)
whitelist = set(whitelist)
features_metadata = features_metadata[features_metadata.rsid.isin(
whitelist)]
logging.info("Opening features")
features = pq.ParquetFile(args.features)
logging.info("Setting R seed")
seed = numpy.random.randint(1e8)
if args.run_tag:
d = pandas.DataFrame({
"run": [args.run_tag],
"cv_seed": [seed]
})[["run", "cv_seed"]]
for t in tissue_names:
Utilities.save_dataframe(
d, "{}_{}_t_runs.txt.gz".format(args.output_prefix, t))
failed_run = False
try:
for i, data_annotation_ in enumerate(data_annotation.itertuples()):
logging.log(9, "processing %i/%i:%s", i + 1,
data_annotation.shape[0], data_annotation_.gene_id)
logging.log(8, "loading data")
d_ = {}
for k, v in data.items():
d_[k] = Parquet._read(v, [data_annotation_.gene_id],
to_pandas=True)
features_ = Genomics.entries_for_gene_annotation(
data_annotation_, args.window, features_metadata)
if features_.shape[0] == 0:
logging.log(9, "No features available")
continue
features_data_ = Parquet._read(features,
[x for x in features_.id.values],
to_pandas=True)
features_data_["id"] = range(1, features_data_.shape[0] + 1)
features_data_ = features_data_[["individual", "id"] +
[x for x in features_.id.values]]
logging.log(8, "training")
prepare_ctimp(args.script_path, seed, args.intermediate_folder,
data_annotation_, features_, features_data_, d_)
del (features_data_)
del (d_)
if args.skip_regression:
continue
subprocess.call([
"bash",
_execution_script(args.intermediate_folder,
data_annotation_.gene_id)
])
w = pandas.read_table(_weights(args.intermediate_folder,
data_annotation_.gene_id),
sep="\s+")
s = pandas.read_table(_summary(args.intermediate_folder,
data_annotation_.gene_id),
sep="\s+")
for e_, entry in enumerate(s.itertuples()):
entry_weights = w[["SNP", "REF.0.", "ALT.1.",
entry.tissue]].rename(
columns={
"SNP": "varID",
"REF.0.": "ref_allele",
"ALT.1.": "eff_allele",
entry.tissue: "weight"
})
entry_weights = entry_weights[entry_weights.weight != 0]
entry_weights = entry_weights.assign(
gene=data_annotation_.gene_id)
entry_weights = entry_weights.merge(features_,
left_on="varID",
right_on="id",
how="left")
entry_weights = entry_weights[WEIGHTS_FIELDS]
if args.output_rsids:
entry_weights.loc[entry_weights.rsid == "NA",
"rsid"] = entry_weights.loc[
entry_weights.rsid == "NA", "varID"]
weights[entry.tissue].write(
entry_weights.to_csv(sep="\t",
index=False,
header=False,
na_rep="NA").encode())
entry_summary = s[s.tissue == entry.tissue].rename(
columns={
"zscore_pval": "pred.perf.pval",
"rho_avg_squared": "pred.perf.R2"
})
entry_summary = entry_summary.assign(
gene=data_annotation_.gene_id,
alpha=0.5,
genename=data_annotation_.gene_name,
gene_type=data_annotation_.gene_type,
n_snps_in_window=features_.shape[0])
entry_summary["n.snps.in.model"] = entry_weights.shape[0]
#must repeat strings beause of weird pandas indexing issue
entry_summary = entry_summary.drop(
["R2", "n", "tissue"], axis=1)[[
"gene", "genename", "gene_type", "alpha",
"n_snps_in_window", "n.snps.in.model", "rho_avg",
"pred.perf.R2", "pred.perf.pval"
]]
summaries[entry.tissue].write(
entry_summary.to_csv(sep="\t",
index=False,
header=False,
na_rep="NA").encode())
features_data_ = Parquet._read(
features, [x for x in entry_weights.varID.values],
to_pandas=True)
var_ids = [x for x in entry_weights.varID.values]
cov = numpy.cov([features_data_[k] for k in var_ids], ddof=1)
ids = [x for x in entry_weights.rsid.values
] if args.output_rsids else var_ids
cov = matrices._flatten_matrix_data([(data_annotation_.gene_id,
ids, cov)])
for cov_ in cov:
l = "{} {} {} {}\n".format(cov_[0], cov_[1], cov_[2],
cov_[3]).encode()
covariances[entry.tissue].write(l)
if not args.keep_intermediate_folder:
logging.info("Cleaning up")
shutil.rmtree(
_intermediate_folder(args.intermediate_folder,
data_annotation_.gene_id))
if args.MAX_M and i >= args.MAX_M:
logging.info("Early abort")
break
except Exception as e:
logging.info("Exception running model training:\n%s",
traceback.format_exc())
failed_run = True
finally:
pass
# This is wrong but I must thing about what to do
#if not args.keep_intermediate_folder:
# shutil.rmtree(args.intermediate_folder)
if not args.skip_regression:
set_down(weights, summaries, covariances, tissue_names, failed_run)
if failed_run:
logging.info("Errors found")
else:
logging.info("Finished")
def run(args):
wp = args.output_prefix + "_weights.txt.gz"
if os.path.exists(wp):
logging.info("Weights output exists already, delete it or move it")
return
sp = args.output_prefix + "_summary.txt.gz"
if os.path.exists(sp):
logging.info("Summary output exists already, delete it or move it")
return
cp = args.output_prefix + "_covariance.txt.gz"
if os.path.exists(wp):
logging.info("covariance output exists already, delete it or move it")
return
r = args.output_prefix + "_run.txt.gz"
if os.path.exists(wp):
logging.info("run output exists already, delete it or move it")
return
logging.info("Starting")
Utilities.ensure_requisite_folders(args.output_prefix)
logging.info("Opening data")
data = pq.ParquetFile(args.data)
available_data = {x for x in data.metadata.schema.names}
logging.info("Loading data annotation")
data_annotation = StudyUtilities.load_gene_annotation(args.data_annotation, args.chromosome, args.sub_batches, args.sub_batch, args.simplify_data_annotation)
data_annotation = data_annotation[data_annotation.gene_id.isin(available_data)]
if args.gene_whitelist:
logging.info("Applying gene whitelist")
data_annotation = data_annotation[data_annotation.gene_id.isin(set(args.gene_whitelist))]
logging.info("Kept %i entries", data_annotation.shape[0])
logging.info("Opening features annotation")
if not args.chromosome:
features_metadata = pq.read_table(args.features_annotation).to_pandas()
else:
features_metadata = pq.ParquetFile(args.features_annotation).read_row_group(args.chromosome-1).to_pandas()
if args.output_rsids:
if not args.keep_highest_frequency_rsid_entry and features_metadata[(features_metadata.rsid != "NA") & features_metadata.rsid.duplicated()].shape[0]:
logging.warning("Several variants map to a same rsid (hint: multiple INDELS?).\n"
"Can't proceed. Consider the using the --keep_highest_frequency_rsid flag, or models will be ill defined.")
return
if args.chromosome and args.sub_batches:
logging.info("Trimming variants")
features_metadata = StudyUtilities.trim_variant_metadata_on_gene_annotation(features_metadata, data_annotation, args.window)
logging.info("Kept %d", features_metadata.shape[0])
if args.variant_call_filter:
logging.info("Filtering variants by average call rate")
features_metadata = features_metadata[features_metadata.avg_call > args.variant_call_filter]
logging.info("Kept %d", features_metadata.shape[0])
if args.variant_r2_filter:
logging.info("Filtering variants by imputation R2")
features_metadata = features_metadata[features_metadata.r2 > args.variant_r2_filter]
logging.info("Kept %d", features_metadata.shape[0])
if args.variant_variance_filter:
logging.info("Filtering variants by (dosage/2)'s variance")
features_metadata = features_metadata[features_metadata["std"]/2 > numpy.sqrt(args.variant_variance_filter)]
logging.info("Kept %d", features_metadata.shape[0])
if args.discard_palindromic_snps:
logging.info("Discarding palindromic snps")
features_metadata = Genomics.discard_gtex_palindromic_variants(features_metadata)
logging.info("Kept %d", features_metadata.shape[0])
if args.rsid_whitelist:
logging.info("Filtering features annotation for whitelist")
whitelist = TextFileTools.load_list(args.rsid_whitelist)
whitelist = set(whitelist)
features_metadata = features_metadata[features_metadata.rsid.isin(whitelist)]
logging.info("Kept %d", features_metadata.shape[0])
if args.only_rsids:
logging.info("discarding non-rsids")
features_metadata = StudyUtilities.trim_variant_metadata_to_rsids_only(features_metadata)
logging.info("Kept %d", features_metadata.shape[0])
if args.keep_highest_frequency_rsid_entry and features_metadata[(features_metadata.rsid != "NA") & features_metadata.rsid.duplicated()].shape[0]:
logging.info("Keeping only the highest frequency entry for every rsid")
k = features_metadata[["rsid", "allele_1_frequency", "id"]]
k.loc[k.allele_1_frequency > 0.5, "allele_1_frequency"] = 1 - k.loc[k.allele_1_frequency > 0.5, "allele_1_frequency"]
k = k.sort_values(by=["rsid", "allele_1_frequency"], ascending=False)
k = k.groupby("rsid").first().reset_index()
features_metadata = features_metadata[features_metadata.id.isin(k.id)]
logging.info("Kept %d", features_metadata.shape[0])
else:
logging.info("rsids are unique, no need to restrict to highest frequency entry")
if args.features_weights:
logging.info("Loading weights")
x_weights = get_weights(args.features_weights, {x for x in features_metadata.id})
logging.info("Filtering features metadata to those available in weights")
features_metadata = features_metadata[features_metadata.id.isin(x_weights.id)]
logging.info("Kept %d entries", features_metadata.shape[0])
else:
x_weights = None
logging.info("Opening features")
features = pq.ParquetFile(args.features)
logging.info("Setting R seed")
s = numpy.random.randint(1e8)
set_seed(s)
if args.run_tag:
d = pandas.DataFrame({"run":[args.run_tag], "cv_seed":[s]})[["run", "cv_seed"]]
Utilities.save_dataframe(d, r)
WEIGHTS_FIELDS=["gene", "rsid", "varID", "ref_allele", "eff_allele", "weight"]
SUMMARY_FIELDS=["gene", "genename", "gene_type", "alpha", "n_snps_in_window", "n.snps.in.model",
"test_R2_avg", "test_R2_sd", "cv_R2_avg", "cv_R2_sd", "in_sample_R2", "nested_cv_fisher_pval",
"nested_cv_converged", "rho_avg", "rho_se", "rho_zscore", "pred.perf.R2", "pred.perf.pval", "pred.perf.qval"]
train = train_elastic_net_wrapper if args.mode == "elastic_net" else train_ols
available_individuals = check_missing(args, data, features)
with gzip.open(wp, "w") as w:
w.write(("\t".join(WEIGHTS_FIELDS) + "\n").encode())
with gzip.open(sp, "w") as s:
s.write(("\t".join(SUMMARY_FIELDS) + "\n").encode())
with gzip.open(cp, "w") as c:
c.write("GENE RSID1 RSID2 VALUE\n".encode())
for i,data_annotation_ in enumerate(data_annotation.itertuples()):
if args.MAX_M and i>=args.MAX_M:
logging.info("Early abort")
break
logging.log(9, "processing %i/%i:%s", i+1, data_annotation.shape[0], data_annotation_.gene_id)
if args.repeat:
for j in range(0, args.repeat):
logging.log(9, "%i-th reiteration", j)
process(w, s, c, data, data_annotation_, features, features_metadata, x_weights, SUMMARY_FIELDS, train, j, nested_folds=args.nested_cv_folds, use_individuals=available_individuals)
else:
process(w, s, c, data, data_annotation_, features, features_metadata, x_weights, SUMMARY_FIELDS, train, nested_folds=args.nested_cv_folds, use_individuals=available_individuals)
logging.info("Finished")
