svincoff's picture
training works
29899b4
Raw
History Blame Contribute Delete
8.8 kB
import pandas as pd
from omegaconf import DictConfig
from pathlib import Path
import os
import rootutils
from dpacman.utils import pylogger
root = rootutils.setup_root(__file__, indicator=".project-root", pythonpath=True)
logger = pylogger.RankedLogger(__name__, rank_zero_only=True)
def clean_nr(nr_raw_path: Path | str):
"""
Clean the non-redundant peaks BED file.
Delete duplicate rows, assign columns, only keep columns we need.
"""
nr = pd.read_csv(nr_raw_path, sep="\t", header=None)
nr.columns = [
"chrom",
"chromStart",
"chromEnd",
"biotypes",
"score",
"strand",
"thickStart",
"thickEnd",
"itemRgb",
]
# make sure we correctly interpret column "biotype" as having one transcription factor separated from all relevant biotypes by ONE colon
biotype_colon_counts = (
nr["biotypes"]
.str.count(":")
.value_counts()
.reset_index()["biotypes"]
.unique()
.tolist()
)
assert biotype_colon_counts == [
1
] # confirm belief that : separates the name of a transcription factor from its biotype - just ONE biotype.
# then split the column accordingly into tr (transcriptional regulator) and biotypes
nr[["tr", "biotypes"]] = nr["biotypes"].str.split(":", expand=True)
# group and concat the scores
logger.info(
f"Keeping only the following columns: chrom, chromStart, chromEnd, biotypes, tr, score."
)
nr = nr[["chrom", "chromStart", "chromEnd", "biotypes", "score", "tr"]]
# drop duplicate rows - all fields
logger.info(f"Size of database before dropping duplicate rows: {len(nr)}")
nr = nr.drop_duplicates().reset_index(drop=True)
logger.info(f"Size of database after dropping duplicate rows: {len(nr)}")
# look for duplicate rows where it's clearly the same experiment but somehow different scores - chrom, chromStart, chromEnd, tr, biotypes
experiment_dups = len(
nr.loc[
nr.duplicated(subset=["chrom", "chromStart", "chromEnd", "tr", "biotypes"])
]
)
logger.info(
f"{experiment_dups} total rows with same chrom, chromStart, chromEnd, biotypes, tr but different score."
)
logger.info(
f"Grouping by everything except score, comma-concatenating unique scores"
)
nr = (
nr.groupby(["chrom", "chromStart", "chromEnd", "tr", "biotypes"])
.agg({"score": lambda x: ",".join(map(str, sorted(set(x))))})
.reset_index()
)
logger.info(f"Final database size: {len(nr)}")
nr["chromLen"] = nr["chromEnd"] - nr["chromStart"]
return nr
def clean_crm(crm_raw_path: Path | str):
"""
Clean the CRM BED file.
Delete duplicate rows, assign columns, only keep columns we need.
"""
crm = pd.read_csv(crm_raw_path, sep="\t", header=None)
crm.columns = [
"chrom",
"chromStart",
"chromEnd",
"tr",
"score",
"strand",
"thickStart",
"thickEnd",
"reserved",
]
# group and concat the scores
logger.info(
f"Keeping only the following columns: chrom, chromStart, chromEnd, tr, score."
)
crm = crm[["chrom", "chromStart", "chromEnd", "tr", "score"]]
# drop duplicate rows - all fields
logger.info(f"Size of database before dropping duplicate rows: {len(crm)}")
crm = crm.drop_duplicates().reset_index(drop=True)
logger.info(f"Size of database after dropping duplicate rows: {len(crm)}")
# look for duplicate rows where it's clearly the same experiment but somehow different scores - chrom, chromStart, chromEnd, tr
experiment_dups = len(
crm.loc[crm.duplicated(subset=["chrom", "chromStart", "chromEnd", "tr"])]
)
logger.info(
f"{experiment_dups} total rows with same chrom, chromStart, chromEnd, tr but different score."
)
logger.info(
f"Grouping by everything except score, comma-concatenating unique scores"
)
crm = (
crm.groupby(["chrom", "chromStart", "chromEnd", "tr"])
.agg({"score": lambda x: ",".join(map(str, sorted(set(x))))})
.reset_index()
)
logger.info(f"Final database size: {len(crm)}")
crm["chromLen"] = crm["chromEnd"] - crm["chromStart"]
return crm
def main(cfg: DictConfig):
# Define the paths
nr_raw_path = Path(root) / cfg.data_task.nr_raw_path
nr_processed_dir = Path(root) / cfg.data_task.nr_processed_dir
nr_processed_filename = cfg.data_task.nr_processed_filename
nr_savepath = os.path.join(nr_processed_dir, nr_processed_filename)
crm_raw_path = Path(root) / cfg.data_task.crm_raw_path
crm_processed_dir = Path(root) / cfg.data_task.crm_processed_dir
crm_processed_filename = cfg.data_task.crm_processed_filename
crm_savepath = os.path.join(crm_processed_dir, crm_processed_filename)
os.makedirs(nr_processed_dir, exist_ok=True)
os.makedirs(crm_processed_dir, exist_ok=True)
# Clean and save the non redundant peaks file
if not (os.path.exists(nr_savepath)):
nr_cleaned = clean_nr(nr_raw_path)
nr_cleaned.to_csv(nr_savepath, sep="\t", index=False)
logger.info(
f"Saved cleaned non-redundant peaks (NR) database to: {nr_savepath}"
)
else:
nr_cleaned = None
logger.info(f"File already exists at {nr_savepath}. Skipping")
# Clean and save the CRM file
if not (os.path.exists(crm_savepath)):
crm_cleaned = clean_crm(crm_raw_path)
crm_cleaned.to_csv(crm_savepath, sep="\t", index=False)
logger.info(
f"Saved cleaned cis-regulatory modules (CRM) database to: {crm_savepath}"
)
else:
crm_cleaned = None
logger.info(f"File already exists at {crm_savepath}. Skipping")
# Save example files
if cfg.data_task.save_example_files:
example_nr_dir = nr_processed_dir / "examples"
os.makedirs(example_nr_dir, exist_ok=True)
example_nr_savepath = os.path.join(
example_nr_dir, "example500_" + nr_processed_filename
)
if not (os.path.exists(example_nr_savepath)):
if nr_cleaned is None:
nr_cleaned = pd.read_csv(nr_savepath, sep="\t")
nr_cleaned.sample(n=500, random_state=42).reset_index(drop=True).to_csv(
example_nr_savepath, sep="\t", index=False
)
logger.info(
f"Saved example NR file with 500 rows to: {example_nr_savepath}"
)
else:
logger.info(
f"Example file already exists at {example_nr_savepath}. Skipping"
)
# CRM example
example_crm_dir = crm_processed_dir / "examples"
os.makedirs(example_crm_dir, exist_ok=True)
example_crm_savepath = os.path.join(
example_crm_dir, "example500_" + crm_processed_filename
)
if not (os.path.exists(example_crm_savepath)):
if crm_cleaned is None:
crm_cleaned = pd.read_csv(crm_savepath, sep="\t")
crm_cleaned.sample(n=500, random_state=42).reset_index(drop=True).to_csv(
example_crm_savepath, sep="\t", index=False
)
logger.info(
f"Saved example CRM file with 500 rows to: {example_crm_savepath}"
)
else:
logger.info(
f"Example file already exists at {example_crm_savepath}. Skipping"
)
# CRM example for one transcription factor
example_crm_tf_savepath = os.path.join(
example_crm_dir, "example500_ERG_" + crm_processed_filename
)
if not (os.path.exists(example_crm_tf_savepath)):
if crm_cleaned is None:
crm_cleaned = pd.read_csv(crm_savepath, sep="\t")
crm_example_tf_db = crm_cleaned.copy(deep=True)
crm_example_tf_db["tr"] = crm_example_tf_db["tr"].apply(
lambda x: x.split(",")
)
crm_example_tf_db = crm_example_tf_db.explode("tr").reset_index(drop=True)
crm_example_tf_db = crm_example_tf_db.loc[crm_example_tf_db["tr"] == "ERG"]
crm_example_tf_db = crm_example_tf_db.sample(
n=min(500, len(crm_example_tf_db)), random_state=42
).reset_index(drop=True)
crm_example_tf_db.to_csv(example_crm_tf_savepath, sep="\t", index=False)
logger.info(
f"Saved example CRM file for one TF with 500 rows to: {example_crm_tf_savepath}"
)
else:
logger.info(
f"Example file already exists at {example_crm_tf_savepath}. Skipping"
)
if __name__ == "__main__":
main()