"""
This file contains all high-level functionality for preprocessing sequencing
data into character matrices ready for phylogenetic inference. This file
is mainly invoked by cassiopeia_preprocess.py.
"""
import warnings
from functools import partial
import os
from pathlib import Path
import time
from typing import List, Optional, Tuple, Union
from Bio import SeqIO
from joblib import delayed
import matplotlib.pyplot as plt
import ngs_tools as ngs
import numpy as np
import pandas as pd
import pysam
from tqdm.auto import tqdm
from typing_extensions import Literal
from cassiopeia.mixins import logger, PreprocessError
from cassiopeia.mixins.warnings import PreprocessWarning
from cassiopeia.preprocess import (
alignment_utilities,
constants,
map_utils,
doublet_utils,
lineage_utils,
UMI_utils,
utilities,
)
DNA_SUBSTITUTION_MATRIX = constants.DNA_SUBSTITUTION_MATRIX
BAM_CONSTANTS = constants.BAM_CONSTANTS
progress = tqdm
[docs]
@logger.namespaced("convert")
@utilities.log_kwargs
@utilities.log_runtime
def convert_fastqs_to_unmapped_bam(
fastq_fps: List[str],
chemistry: Literal["dropseq", "10xv2", "10xv3", "indropsv3", "slideseq2"],
output_directory: str,
name: Optional[str] = None,
n_threads: int = 1,
) -> str:
"""Converts FASTQs into an unmapped BAM based on a chemistry.
This function converts a set of FASTQs into an unmapped BAM with appropriate
BAM tags.
Args:
fastq_fps: List of paths to FASTQ files. Usually, this argument contains
two FASTQs, where the first contains the barcode and UMI sequences
and the second contains cDNA. The FASTQs may be gzipped.
chemistry: Sample-prep/sequencing chemistry used. The following
chemistries are supported:
* dropseq: Droplet-based scRNA-seq chemistry described in
Macosco et al. 2015
* 10xv2: 10x Genomics 3' version 2
* 10xv3: 10x Genomics 3' version 3
* indropsv3: inDrops version 3 by Zilionis et al. 2017
* slideseq2: Slide-seq version 2
output_directory: The output directory where the unmapped BAM will be
written to. This directory must exist prior to calling this function.
name: Name of the reads in the FASTQs. This name is set as the read group
name for the reads in the output BAM, as well as the filename prefix
of the output BAM. If not provided, a short random UUID is used as
the read group name, but not as the filename prefix of the BAM.
n_threads: Number of threads to use. Defaults to 1.
Returns:
Path to written BAM
Raises:
PreprocessError if the provided chemistry does not exist.
"""
if chemistry not in constants.CHEMISTRY_BAM_TAGS:
raise PreprocessError(f"Unknown chemistry {chemistry}")
tag_map = constants.CHEMISTRY_BAM_TAGS[chemistry]
chemistry = ngs.chemistry.get_chemistry(chemistry)
logger.info(f"Using {chemistry} chemistry.")
bam_fp = os.path.join(
output_directory, f"{name}_unmapped.bam" if name else "unmapped.bam"
)
ngs.fastq.fastqs_to_bam_with_chemistry(
fastq_fps,
chemistry,
tag_map,
bam_fp,
name=name,
show_progress=True,
n_threads=n_threads,
)
return bam_fp
[docs]
@logger.namespaced("filter_bam")
@utilities.log_kwargs
@utilities.log_runtime
def filter_bam(
bam_fp: str,
output_directory: str,
quality_threshold: int = 10,
n_threads: int = 1,
) -> str:
"""Filter reads in a BAM that have low quality barcode or UMIs.
Args:
bam_fp: Input BAM filepath containing reads to filter.
output_directory: The output directory where the filtered BAM will be
written to. This directory must exist prior to calling this function.
quality_threshold: Every base of the barcode and UMI sequence for a
given read must have at least this PHRED quality score for it to
pass the filtering.
n_threads: Number of threads to use. Defaults to 1.
Returns:
Path to filtered BAM
"""
n_filtered = 0
def filter_func(aln):
# False means this read will be filtered out
filter_bool = all(
q >= quality_threshold
for q in pysam.qualitystring_to_array(
aln.get_tag(BAM_CONSTANTS["RAW_CELL_BC_QUALITY_TAG"])
)
) and all(
q >= quality_threshold
for q in pysam.qualitystring_to_array(
aln.get_tag(BAM_CONSTANTS["UMI_QUALITY_TAG"])
)
)
nonlocal n_filtered
n_filtered += not filter_bool
return filter_bool
prefix, ext = os.path.splitext(os.path.basename(bam_fp))
filtered_fp = os.path.join(output_directory, f"{prefix}_filtered{ext}")
ngs.bam.filter_bam(
bam_fp,
filter_func,
filtered_fp,
show_progress=True,
n_threads=n_threads,
)
logger.info(f"Filtered {n_filtered} reads that didn't pass the filter.")
return filtered_fp
[docs]
@logger.namespaced("error_correct_cellbcs_to_whitelist")
@utilities.log_kwargs
@utilities.log_runtime
def error_correct_cellbcs_to_whitelist(
bam_fp: str,
whitelist: Union[str, List[str]],
output_directory: str,
n_threads: int = 1,
) -> str:
"""Error-correct cell barcodes in the input BAM.
The barcode correction procedure used in Cell Ranger by 10X Genomics is used.
https://kb.10xgenomics.com/hc/en-us/articles/115003822406-How-does-Cell-Ranger-correct-barcode-sequencing-errors
This function can either take a list of whitelisted barcodes or a plaintext
file containing these barcodes.
Args:
bam_fp: Input BAM filepath containing raw barcodes
whitelist: May be either a single path to a plaintext file containing
the barcode whitelist, one barcode per line, or a list of
barcodes.
output_directory: The output directory where the corrected BAM will be
written to. This directory must exist prior to calling this function.
n_threads: Number of threads to use. Defaults to 1.
Todo:
Currently, the user must provide their own whitelist, and Cassiopeia
does not use any of the whitelists provided by the ngs-tools library.
At some point, we should update the pipeline so that if ngs-tools
provides a pre-packaged whitelists, it uses that for those chemistries.
Returns:
Path to corrected BAM
"""
if isinstance(whitelist, list):
whitelist_set = set(whitelist)
else:
with open(whitelist, "r") as f:
whitelist_set = set(
line.strip() for line in f if not line.isspace()
)
whitelist = list(whitelist_set)
# Extract all raw barcodes and their qualities
barcodes = []
qualities = []
with pysam.AlignmentFile(
bam_fp, "rb", check_sq=False, threads=n_threads
) as f:
for read in f:
barcodes.append(read.get_tag(BAM_CONSTANTS["RAW_CELL_BC_TAG"]))
qualities.append(
read.get_tag(BAM_CONSTANTS["RAW_CELL_BC_QUALITY_TAG"])
)
logger.info(f"Detected {len(set(barcodes))} raw barcodes.")
# Correct
corrections = ngs.sequence.correct_sequences_to_whitelist(
barcodes, qualities, whitelist, show_progress=True, n_threads=n_threads
)
# Write corrected BAM
prefix, ext = os.path.splitext(os.path.basename(bam_fp))
corrected_fp = os.path.join(output_directory, f"{prefix}_corrected{ext}")
with pysam.AlignmentFile(
bam_fp, "rb", check_sq=False, threads=n_threads
) as f_in:
with pysam.AlignmentFile(
corrected_fp, "wb", template=f_in, threads=n_threads
) as f_out:
for i, read in enumerate(f_in):
if corrections[i]:
read.set_tag(BAM_CONSTANTS["CELL_BC_TAG"], corrections[i])
f_out.write(read)
return corrected_fp
[docs]
@logger.namespaced("collapse")
@utilities.log_kwargs
@utilities.log_runtime
def collapse_umis(
bam_fp: str,
output_directory: str,
max_hq_mismatches: int = 3,
max_indels: int = 2,
method: Literal["cutoff", "likelihood"] = "cutoff",
n_threads: int = 1,
) -> pd.DataFrame:
"""Collapses close UMIs together from a bam file.
On a basic level, it aggregates together identical or close reads to count
how many times a UMI was read. Performs basic error correction, allowing
UMIs to be collapsed together which differ by at most a certain number of
high quality mismatches and indels in the sequence read itself. Writes out
a dataframe of the collapsed UMIs table.
Args:
bam_file_name: File path of the bam_file. Just the bam file name can be
specified if the bam already exists in the output directory
output_directory: The output directory where the sorted bam directory, the
collapsed bam directory, and the final collapsed table are written to
max_hq_mismatches: A threshold specifying the max number of high quality
mismatches between the seqeunces of 2 aligned segments to be collapsed
max_indels: A threshold specifying the maximum number of differing indels
allowed between the sequences of 2 aligned segments to be collapsed
method: Which method to use to form initial sequence clusters. Must be
one of the following:
* cutoff: Uses a quality score hard cutoff of 30, and any mismatches
below this quality are ignored. Initial sequence clusters are
formed by selecting the most common base at each position (with
quality at least 30).
* likelihood: Utilizes the error probability encoded in the quality
score. Initial sequence clusters are formed by selecting the
most probable at each position.
n_threads: Number of threads to use.
Returns:
A DataFrame of collapsed reads.
"""
# pathing written such that the bam file that is being converted does not
# have to exist currently in the output directory
if output_directory[-1] == "/":
output_directory = output_directory[:-1]
sorted_file_name = Path(
output_directory
+ "/"
+ ".".join(bam_fp.split("/")[-1].split(".")[:-1])
+ "_sorted.bam"
)
cell_bc_tag = UMI_utils.detect_cell_bc_tag(bam_fp)
logger.info(f"Using BAM tag `{cell_bc_tag}` as cell barcodes")
max_read_length, total_reads_out = UMI_utils.sort_bam(
bam_fp,
str(sorted_file_name),
sort_key=lambda al: (
al.get_tag(cell_bc_tag),
al.get_tag(BAM_CONSTANTS["UMI_TAG"]),
),
filter_func=lambda al: al.has_tag(cell_bc_tag),
)
logger.info("Sorted bam directory saved to " + str(sorted_file_name))
logger.info("Max read length of " + str(max_read_length))
logger.info("Total reads: " + str(total_reads_out))
collapsed_file_name = sorted_file_name.with_suffix(".collapsed.bam")
UMI_utils.form_collapsed_clusters(
str(sorted_file_name),
str(collapsed_file_name),
max_hq_mismatches,
max_indels,
cell_key=lambda al: al.get_tag(cell_bc_tag),
method=method,
n_threads=n_threads,
)
collapsed_df_file_name = sorted_file_name.with_suffix(".collapsed.txt")
df = utilities.convert_bam_to_df(str(collapsed_file_name))
df.to_csv(str(collapsed_df_file_name), sep="\t", index=False)
logger.info("Collapsed bam directory saved to " + str(collapsed_file_name))
logger.info("Converted dataframe saved to " + str(collapsed_df_file_name))
return df
[docs]
@logger.namespaced("resolve")
@utilities.log_kwargs
@utilities.log_runtime
def resolve_umi_sequence(
molecule_table: pd.DataFrame,
output_directory: str,
min_umi_per_cell: int = 10,
min_avg_reads_per_umi: float = 2.0,
plot: bool = True,
) -> pd.DataFrame:
"""Resolve a consensus sequence for each UMI.
This procedure will perform UMI and cellBC filtering on the basis of reads
per UMI and UMIs per cell and then assign the most abundant sequence to
each UMI if there is a set of conflicting sequences per UMI.
Args:
molecule_table: molecule table to resolve
output_directory: Directory to store results
min_umi_per_cell: The threshold specifying the minimum number of UMIs
in a cell needed to be retained during filtering
min_avg_reads_per_umi: The threshold specifying the minimum coverage
(i.e. average) reads per UMI in a cell needed for that cell to be
retained during filtering
Returns:
A molecule table with unique mappings between cellBC-UMI pairs.
"""
if plot:
# -------------------- Plot # of sequences per UMI -------------------- #
equivClass_group = (
molecule_table.groupby(["cellBC", "UMI"])
.agg({"grpFlag": "count"})
.sort_values("grpFlag", ascending=False)
.reset_index()
)
_ = plt.figure(figsize=(8, 5))
plt.hist(
equivClass_group["grpFlag"],
bins=range(1, equivClass_group["grpFlag"].max()),
)
plt.title("Unique Seqs per cellBC+UMI")
plt.yscale("log", base=10)
plt.xlabel("Number of Unique Seqs")
plt.ylabel("Count (Log)")
plt.savefig(os.path.join(output_directory, "seqs_per_equivClass.png"))
plt.close()
# ----------------- Select most abundant sequence ------------------ #
mt_filter = {}
total_numReads = {}
top_reads = {}
second_reads = {}
first_reads = {}
unique_pairs = molecule_table.groupby(["cellBC", "UMI"], sort=False)
for _, group in tqdm(
unique_pairs,
total=len(unique_pairs.size()),
desc="Resolving UMI sequences",
):
# base case - only one sequence
if group.shape[0] == 1:
good_readName = group["readName"].iloc[0]
mt_filter[good_readName] = False
total_numReads[good_readName] = group["readCount"].values[0]
top_reads[good_readName] = group["readCount"].values[0]
# more commonly - many sequences for a given UMI
else:
group_sort = group.sort_values(
"readCount", ascending=False, ignore_index=True
)
good_readName = group_sort["readName"].iloc[0]
# keep the first entry (highest readCount)
mt_filter[good_readName] = False
total_numReads[good_readName] = group_sort["readCount"].sum()
top_reads[good_readName] = group_sort["readCount"].iloc[0]
second_reads[good_readName] = group_sort["readCount"].iloc[1]
first_reads[good_readName] = group_sort["readCount"].iloc[0]
# mark remaining UMIs for filtering
for i in range(1, group.shape[0]):
bad_readName = group_sort["readName"].iloc[i]
mt_filter[bad_readName] = True
# apply the filter using the hash table created above
molecule_table["filter"] = molecule_table["readName"].map(mt_filter)
n_filtered = molecule_table[molecule_table["filter"]].shape[0]
logger.info(f"Filtered out {n_filtered} reads.")
# filter based on status & reindex
filt_molecule_table = molecule_table[~molecule_table["filter"]].copy()
filt_molecule_table.drop(columns=["filter"], inplace=True)
if plot:
# ---------------- Plot Diagnositics after Resolving ---------------- #
h = plt.figure(figsize=(14, 10))
plt.plot(list(top_reads.values()), list(total_numReads.values()), "r.")
plt.ylabel("Total Reads")
plt.xlabel("Number Reads for Picked Sequence")
plt.title("Total vs. Top Reads for Picked Sequence")
plt.savefig(
os.path.join(output_directory, "total_vs_top_reads_pickSeq.png")
)
plt.close()
h = plt.figure(figsize=(14, 10))
plt.plot(list(first_reads.values()), list(second_reads.values()), "r.")
plt.ylabel("Number Reads for Second Best Sequence")
plt.xlabel("Number Reads for Picked Sequence")
plt.title("Second Best vs. Top Reads for Picked Sequence")
plt.savefig(
os.path.join(output_directory, "second_vs_top_reads_pickSeq.png")
)
plt.close()
filt_molecule_table = utilities.filter_cells(
filt_molecule_table,
min_umi_per_cell=min_umi_per_cell,
min_avg_reads_per_umi=min_avg_reads_per_umi,
)
return filt_molecule_table
[docs]
@logger.namespaced("align")
@utilities.log_kwargs
@utilities.log_runtime
def align_sequences(
queries: pd.DataFrame,
ref_filepath: Optional[str] = None,
ref: Optional[str] = None,
gap_open_penalty: float = 20,
gap_extend_penalty: float = 1,
method: Literal["local", "global"] = "local",
n_threads: int = 1,
) -> pd.DataFrame:
"""Align reads to the TargetSite reference.
Take in several queries stored in a DataFrame mapping cellBC-UMIs to a
sequence of interest and align each to a reference sequence. Either local
or global alignment may be performed, depending on the `method` argument.
The defaults for the gap open and gap extend penalties were selected via
in-silico simulation (and are functionally equivalent to the values used
in the GESTALT technology described in McKenna et al, 2016). The desired
output consists of the best alignment score and the CIGAR string storing the
indel locations in the query sequence.
Args:
queries: DataFrame storing a list of sequences to align.
ref_filepath: Filepath to the reference FASTA.
ref: Reference sequence.
gap_open_penalty: Gap open penalty
gap_extend_penalty: Gap extension penalty
method: What alignment algorithm to use. Can be either "local" to
perform local alignment using Smith-Waterman or "global" to
perform global alignment using Needleman Wunsch.
n_threads: Number of threads to use.
Returns:
A DataFrame mapping each sequence name to the CIGAR string, quality,
and original query sequence.
Raises:
PreprocessError if both or neither `ref_filepath` and `ref` are
provided, or if the `method` is not either "local" or "global".
"""
if (ref is None) == (ref_filepath is None):
raise PreprocessError(
"Either `ref_filepath` or `ref` must be provided."
)
if method == "local":
align = alignment_utilities.align_local
elif method == "global":
align = alignment_utilities.align_global
else:
raise PreprocessError("`method` must be either 'local' or 'global'.")
alignment_dictionary = {}
if ref_filepath:
ref = str(list(SeqIO.parse(ref_filepath, "fasta"))[0].seq)
# First, align all unique sequences and save results into dictionary.
align_partial = partial(
align,
substitution_matrix=DNA_SUBSTITUTION_MATRIX,
gap_open_penalty=gap_open_penalty,
gap_extend_penalty=gap_extend_penalty,
)
all_sequences = list(set(queries["seq"]))
alignments = []
for seq, aln in zip(
all_sequences,
ngs.utils.ParallelWithProgress(
n_jobs=n_threads,
total=len(all_sequences),
desc="Aligning sequences to reference",
)(delayed(align_partial)(ref, seq) for seq in all_sequences),
):
alignments.append(aln)
alignment_table = pd.DataFrame(
alignments,
columns=[
"CIGAR",
"QueryBegin",
"ReferenceBegin",
"AlignmentScore",
"seq",
],
)
# Merge alignments into input dataframe
alignment_df = pd.merge(
queries[["readName", "cellBC", "UMI", "readCount", "seq"]],
alignment_table,
how="left",
on="seq",
)
alignment_df.rename(columns={"seq": "Seq"}, inplace=True)
return alignment_df
[docs]
@logger.namespaced("call_alleles")
@utilities.log_kwargs
@utilities.log_runtime
def call_alleles(
alignments: pd.DataFrame,
ref_filepath: Optional[str] = None,
ref: Optional[str] = None,
barcode_interval: Tuple[int, int] = (20, 34),
cutsite_locations: List[int] = [112, 166, 220],
cutsite_width: int = 12,
context: bool = True,
context_size: int = 5,
) -> pd.DataFrame:
"""Call indels from CIGAR strings.
Given many alignments, we extract the indels by comparing the CIGAR strings
of each alignment to the reference sequence.
Args:
alignments: Alignments provided in DataFrame
ref_filepath: Filepath to the reference sequence
ref: Nucleotide sequence of the reference
barcode_interval: Interval in reference corresponding to the integration
barcode
cutsite_locations: A list of all cutsite positions in the reference
cutsite_width: Number of nucleotides left and right of cutsite location
that indels can appear in.
context: Include sequence context around indels
context_size: Number of bases to the right and left to include as
context
Returns:
A DataFrame mapping each sequence alignment to the called indels.
"""
if (ref is None) == (ref_filepath is None):
raise PreprocessError(
"Either `ref_filepath` or `ref` must be provided."
)
alignment_to_indel = {}
alignment_to_intBC = {}
if ref_filepath:
ref = str(list(SeqIO.parse(ref_filepath, "fasta"))[0].seq)
for _, row in tqdm(
alignments.iterrows(),
total=alignments.shape[0],
desc="Parsing CIGAR strings into indels",
):
intBC, indels = alignment_utilities.parse_cigar(
row.CIGAR,
row.Seq,
ref,
row.ReferenceBegin,
row.QueryBegin,
barcode_interval,
cutsite_locations,
cutsite_width,
context=context,
context_size=context_size,
)
alignment_to_indel[row.readName] = indels
alignment_to_intBC[row.readName] = intBC
indel_df = pd.DataFrame.from_dict(
alignment_to_indel,
orient="index",
columns=[f"r{i}" for i in range(1, len(cutsite_locations) + 1)],
)
indel_df["allele"] = indel_df.apply(
lambda x: "".join([str(i) for i in x.values]), axis=1
)
indel_df["intBC"] = indel_df.index.map(alignment_to_intBC)
alignments.set_index("readName", inplace=True)
alignments = alignments.join(indel_df)
alignments.reset_index(inplace=True)
# check cut-sites and raise a warning if any missing data is detected
cutsites = utilities.get_default_cut_site_columns(alignments)
if np.any((alignments[cutsites] == "").sum(axis=0) > 0):
warnings.warn(
"Detected missing data in alleles. You might"
" consider re-running align_sequences with a"
" lower gap-open penalty, or using a separate"
" alignment strategy.",
PreprocessWarning,
)
return alignments
[docs]
@logger.namespaced("error_correct_intbcs_to_whitelist")
@utilities.log_kwargs
@utilities.log_runtime
def error_correct_intbcs_to_whitelist(
input_df: pd.DataFrame,
whitelist: Union[str, List[str]],
intbc_dist_thresh: int = 1,
) -> pd.DataFrame:
"""Corrects all intBCs to the provided whitelist.
This function can either take a list of whitelisted intBCs or a plaintext
file containing these intBCs.
Args:
input_df: Input DataFrame of alignments.
whitelist: May be either a single path to a plaintext file containing
the barcode whitelist, one barcode per line, or a list of
barcodes.
intbc_dist_thresh: The threshold specifying the maximum Levenshtein
distance between the read sequence and whitelist to be corrected.
Returns:
A DataFrame of error corrected intBCs.
"""
if isinstance(whitelist, list):
whitelist_set = set(whitelist)
else:
with open(whitelist, "r") as f:
whitelist_set = set(
line.strip() for line in f if not line.isspace()
)
whitelist = list(whitelist_set)
unique_intbcs = list(input_df["intBC"].unique())
corrections = {intbc: intbc for intbc in whitelist_set}
logger.info(f"{len(unique_intbcs)} intBCs detected.")
for intbc in progress(unique_intbcs, desc="Correcting intBCs to whitelist"):
min_distance = np.inf
min_wls = []
if intbc not in whitelist_set:
for wl_intbc in whitelist:
distance = ngs.sequence.levenshtein_distance(intbc, wl_intbc)
if distance < min_distance:
min_distance = distance
min_wls = [wl_intbc]
elif distance == min_distance:
min_wls.append(wl_intbc)
# Correct only if there is one matching whitelist. Discard if there
# are multiple possible corrections.
if len(min_wls) == 1 and min_distance <= intbc_dist_thresh:
corrections[intbc] = min_wls[0]
input_df["intBC"] = input_df["intBC"].map(corrections)
return input_df[~input_df["intBC"].isna()]
[docs]
@logger.namespaced("error_correct_umis")
@utilities.log_kwargs
@utilities.log_runtime
def error_correct_umis(
input_df: pd.DataFrame,
max_umi_distance: int = 2,
allow_allele_conflicts: bool = False,
n_threads: int = 1,
) -> pd.DataFrame:
"""Within cellBC-intBC pairs, collapses UMIs that have close sequences.
Error correct UMIs together within cellBC-intBC pairs. UMIs that have a
Hamming Distance between their sequences less than a threshold are
corrected towards whichever UMI is more abundant. The `allow_allele_conflicts`
option may be used to also group on the actual allele.
Args:
input_df: Input DataFrame of alignments.
max_umi_distance: The threshold specifying the Maximum Hamming distance
between UMIs for one to be corrected to another.
allow_allele_conflicts: Whether or not to include the allele when
splitting UMIs into allele groups. When True, UMIs are grouped by
cellBC-intBC-allele triplets. When False, UMIs are grouped by
cellBC-intBC pairs. This option is used when it is possible for
each cellBC-intBC pair to have >1 allele state, such as for
spatial data.
n_threads: Number of threads to use.
Returns:
A DataFrame of error corrected UMIs.
"""
if (
len(
[
i
for i in input_df.groupby(["cellBC", "intBC", "UMI"]).size()
if i > 1
]
)
!= 0
):
raise PreprocessError(
"Non-unique cellBC-UMI pair exists, please resolve UMIs."
)
sorted_df = input_df.sort_values(
["cellBC", "intBC", "readCount", "UMI"],
ascending=[True, True, False, False],
)
if max_umi_distance == 0:
logger.info(
"Distance of 0, no correction occurred, all alignments returned"
)
return sorted_df
num_corrected = 0
total = sorted_df.shape[0]
alignment_df = pd.DataFrame()
groupby = ["cellBC", "intBC"]
if allow_allele_conflicts:
groupby.append("allele")
allele_groups = sorted_df.groupby(groupby)
alignment_dfs = []
for allele_group, num_corr in ngs.utils.ParallelWithProgress(
n_jobs=n_threads, total=len(allele_groups), desc="Error-correcting UMIs"
)(
delayed(UMI_utils.correct_umis_in_group)(allele_group, max_umi_distance)
for _, allele_group in allele_groups
):
num_corrected += num_corr
alignment_dfs.append(allele_group)
alignment_df = pd.concat(alignment_dfs, sort=True)
logger.info(
f"{num_corrected} UMIs Corrected of {total}"
+ f"({round(float(num_corrected) / total, 5) * 100}%)"
)
alignment_df["readName"] = alignment_df.apply(
lambda x: "_".join([x.cellBC, x.UMI, str(int(x.readCount))]), axis=1
)
alignment_df.set_index("readName", inplace=True)
alignment_df.reset_index(inplace=True)
return alignment_df
[docs]
@logger.namespaced("filter_molecule_table")
@utilities.log_kwargs
@utilities.log_runtime
def filter_molecule_table(
input_df: pd.DataFrame,
output_directory: str,
min_umi_per_cell: int = 10,
min_avg_reads_per_umi: float = 2.0,
min_reads_per_umi: int = -1,
intbc_prop_thresh: float = 0.5,
intbc_umi_thresh: int = 10,
intbc_dist_thresh: int = 1,
doublet_threshold: float = 0.35,
allow_allele_conflicts: bool = False,
plot: bool = False,
) -> pd.DataFrame:
"""Filters and corrects a molecule table of cellBC-UMI pairs.
Performs the following steps on the alignments in a DataFrame:
1. Filters out UMIs with read count < `min_reads_per_umi`. If
`min_reads_per_umi` is less than 0, a dynamic threshold is
calculated as `(99th percentile of read counts) // 10`.
2. Filters out cellBCs with unique UMIs < `min_umi_per_cell` and
average read count per UMI < `min_avg_reads_per_umi`.
3. Error corrects intBCs by changing intBCs with low UMI counts to
intBCs with the same allele and a close sequence
4. Filters out cellBCs that contain too much conflicting allele
information as intra-lineage doublets
5. Chooses one allele for each cellBC-intBC pair, by selecting the most
common. This is not performed when `allow_allele_conflicts` is True.
Args:
input_df: A molecule table, i.e. cellBC-UMI pairs. Note that
each cellBC should only contain one instance of each UMI
output_directory: The output directory path to store plots
min_umi_per_cell: The threshold specifying the minimum number of UMIs
in a cell needed to be retained during filtering
min_avg_reads_per_umi: The threshold specifying the minimum coverage
(i.e. average) reads per UMI in a cell needed in order for that
cell to be retained during filtering
min_reads_per_umi: The threshold specifying the minimum read count
needed for a UMI to be retained during filtering. Set dynamically
if value is < 0.
intbc_prop_thresh: The threshold specifying the maximum proportion of
the total UMI counts for a intBC to be corrected to another
intbc_umi_thresh: The threshold specifying the maximum UMI count for
an intBC needs to be corrected to another
intbc_dist_thresh: The threshold specifying the maximum Levenshtein
Distance between sequences for an intBC to be corrected to another
doublet_threshold: The threshold specifying the maximum proportion of
conflicting alleles information allowed to for an intBC to be
retained in doublet filtering. Set to None to skip doublet filtering
allow_allele_conflicts: Whether or not to allow multiple alleles to be
assigned to each cellBC-intBC pair. For fully single-cell data,
this option should be set to False, since each cell is expected to
have a single allele state for each intBC. However, this option
should be set to True for chemistries that may result in multiple
physical cells being captured for each barcode.
plot: Indicates whether to plot the change in intBC and cellBC counts
across filtering stages
Returns:
A filtered and corrected allele table of cellBC-UMI-allele groups
"""
input_df["status"] = "good"
input_df.sort_values("readCount", ascending=False, inplace=True)
rc_profile, upi_profile, upc_profile = {}, {}, {}
logger.info("Logging initial stats...")
if plot:
(
rc_profile["Init"],
upi_profile["Init"],
upc_profile["Init"],
) = utilities.record_stats(input_df)
if min_reads_per_umi < 0:
R = input_df["readCount"]
if list(R):
min_reads_per_umi = np.percentile(R, 99) // 10
else:
min_reads_per_umi = 0
logger.info(f"Filtering UMIs with less than {min_reads_per_umi} reads...")
filtered_df = utilities.filter_umis(
input_df, min_reads_per_umi=min_reads_per_umi
)
if plot:
(
rc_profile["Filtered_UMI"],
upi_profile["Filtered_UMI"],
upc_profile["Filtered_UMI"],
) = utilities.record_stats(filtered_df)
logger.info(
f"Filtering out cellBCs with fewer than {min_umi_per_cell} UMIs and"
f"less than {min_avg_reads_per_umi} average reads per UMI..."
)
filtered_df = utilities.filter_cells(
filtered_df,
min_umi_per_cell=min_umi_per_cell,
min_avg_reads_per_umi=min_avg_reads_per_umi,
)
if plot:
(
rc_profile["CellFilter"],
upi_profile["CellFilter"],
upc_profile["CellFilter"],
) = utilities.record_stats(filtered_df)
if intbc_dist_thresh > 0:
logger.info("Error correcting intBCs...")
filtered_df = utilities.error_correct_intbc(
filtered_df,
prop=intbc_prop_thresh,
umi_count_thresh=intbc_umi_thresh,
dist_thresh=intbc_dist_thresh,
)
if plot:
(
rc_profile["Process_intBC"],
upi_profile["Process_intBC"],
upc_profile["Process_intBC"],
) = utilities.record_stats(filtered_df)
if doublet_threshold and not allow_allele_conflicts:
logger.info(
f"Filtering out intra-lineage group doublets with proportion {doublet_threshold}..."
)
filtered_df = doublet_utils.filter_intra_doublets(
filtered_df, prop=doublet_threshold
)
if not allow_allele_conflicts:
logger.info("Mapping remaining intBC conflicts...")
filtered_df = map_utils.map_intbcs(filtered_df)
if plot:
(
rc_profile["Final"],
upi_profile["Final"],
upc_profile["Final"],
) = utilities.record_stats(filtered_df)
# Count total filtered cellBCs
cellBC_count = 0
for name, grp in filtered_df.groupby(["cellBC"]):
cellBC_count += 1
if plot:
stages = [
"Init",
"CellFilter",
"Filtered_UMI",
"Process_intBC",
"Final",
]
# Plot Read Per UMI Histogram
h = plt.figure(figsize=(14, 10))
for n in stages:
ax = plt.hist(
rc_profile[n], label=n, histtype="step", log=True, bins=200
)
plt.legend()
plt.ylabel("Frequency")
plt.xlabel("Number of Reads")
plt.title("Reads Per UMI")
plt.savefig(os.path.join(output_directory, "reads_per_umi.png"))
plt.close()
h = plt.figure(figsize=(14, 10))
for n in stages:
ax = plt.plot(upc_profile[n], label=n)
plt.legend()
plt.ylabel("Number of UMIs")
plt.xlabel("Rank Order")
plt.xscale("log", base=10)
plt.yscale("log", base=10)
plt.title("UMIs per CellBC")
plt.savefig(os.path.join(output_directory, "umis_per_cellbc.png"))
plt.close()
h = plt.figure(figsize=(14, 10))
for n in stages:
ax = plt.hist(
upi_profile[n], label=n, histtype="step", log=True, bins=200
)
plt.legend()
plt.ylabel("Frequency")
plt.xlabel("Number of UMIs")
plt.title("UMIs per intBC")
plt.savefig(os.path.join(output_directory, "umis_per_intbc.png"))
plt.close()
logger.info(
f"Overall, filtered {cellBC_count} cells, with {filtered_df.shape[0]} UMIs."
)
filtered_df.set_index("readName", inplace=True)
filtered_df.reset_index(inplace=True)
return filtered_df
[docs]
@logger.namespaced("call_lineages")
@utilities.log_kwargs
@utilities.log_runtime
def call_lineage_groups(
input_df: pd.DataFrame,
output_directory: str,
min_umi_per_cell: int = 10,
min_avg_reads_per_umi: float = 2.0,
min_cluster_prop: float = 0.005,
min_intbc_thresh: float = 0.05,
inter_doublet_threshold: float = 0.35,
kinship_thresh: float = 0.25,
plot: bool = False,
) -> pd.DataFrame:
"""Assigns cells to their clonal populations.
Performs multiple rounds of filtering and assigning to lineage groups:
1. Iteratively generates putative lineage groups by forming intBC
groups for each lineage group and then assigning cells based on how
many intBCs they share with each intBC group (kinship).
2. Refines these putative groups by removing non-informative intBCs
and reassigning cells through kinship.
3. Removes all inter-lineage doublets, defined as cells that have
relatively equal kinship scores across multiple lineages and whose
assignments are therefore ambigious.
4. Finally, performs one more round of filtering non-informative intBCs
and cellBCs with low UMI counts before returning a final table of
lineage assignments, allele information, and read and umi counts for
each sample.
Args:
input_df: The allele table of cellBC-UMI-allele groups to be annotated
with lineage assignments
output_directory: The folder to store the final table as well as plots
min_umi_per_cell: The threshold specifying the minimum number of UMIs a
cell needs in order to not be filtered during filtering
min_avg_reads_per_umi: The threshold specifying the minimum coverage
(i.e. average) reads per UMI in a cell needed in order for that
cell not to be filtered during filtering
min_cluster_prop: The minimum cluster size in the putative lineage
assignment step, as a proportion of the number of cells
min_intbc_thresh: The threshold specifying the minimum proportion of
cells in a lineage group that need to have an intBC in order for it
be retained during filtering. Also specifies the minimum proportion
of cells that share an intBC with the most frequent intBC in
forming putative lineage groups
inter_doublet_threshold: The threshold specifying the minimum proportion
of kinship a cell shares with its assigned lineage group out of all
lineage groups for it to be retained during doublet filtering
kinship_thresh: The threshold specifying the minimum proportion of
intBCs shared between a cell and the intBC set of a lineage group
needed to assign that cell to that lineage group in putative
assignment
plot: Indicates whether to generate plots
Returns:
None, saves output allele table to file.
"""
logger.info(
f"{input_df.shape[0]} UMIs (rows), with {input_df.shape[1]} attributes (columns)"
)
logger.info(str(len(input_df["cellBC"].unique())) + " Cells")
# Create a pivot_table
piv = pd.pivot_table(
input_df, index="cellBC", columns="intBC", values="UMI", aggfunc="count"
)
piv = piv.div(piv.sum(axis=1), axis=0)
# Reorder piv columns by binarized intBC frequency
pivbin = piv.copy()
pivbin[pivbin > 0] = 1
intBC_sums = pivbin.sum(0)
ordered_intBCs = intBC_sums.sort_values(ascending=False).index.tolist()
piv = piv[ordered_intBCs]
min_clust_size = int(min_cluster_prop * piv.shape[0])
logger.info("Assigning initial lineage groups...")
logger.info(f"Clustering with minimum cluster size {min_clust_size}...")
piv_assigned = lineage_utils.assign_lineage_groups(
piv,
min_clust_size,
min_intbc_thresh=min_intbc_thresh,
kinship_thresh=kinship_thresh,
)
logger.info("Refining lineage groups...")
logger.info(
"Redefining lineage groups by removing low proportion intBCs..."
)
master_LGs, master_intBCs = lineage_utils.filter_intbcs_lg_sets(
piv_assigned, min_intbc_thresh=min_intbc_thresh
)
logger.info("Reassigning cells to refined lineage groups by kinship...")
kinship_scores = lineage_utils.score_lineage_kinships(
piv_assigned, master_LGs, master_intBCs
)
logger.info("Annotating alignment table with refined lineage groups...")
allele_table = lineage_utils.annotate_lineage_groups(
input_df, kinship_scores, master_intBCs
)
if inter_doublet_threshold:
logger.info(
f"Filtering out inter-lineage group doublets with proportion {inter_doublet_threshold}..."
)
allele_table = doublet_utils.filter_inter_doublets(
allele_table, rule=inter_doublet_threshold
)
logger.info(
"Filtering out low proportion intBCs in finalized lineage groups..."
)
filtered_lgs = lineage_utils.filter_intbcs_final_lineages(
allele_table, min_intbc_thresh=min_intbc_thresh
)
allele_table = lineage_utils.filtered_lineage_group_to_allele_table(
filtered_lgs
)
logger.debug("Final lineage group assignments:")
for n, g in allele_table.groupby(["lineageGrp"]):
logger.debug(f"LG {n}: " + str(len(g["cellBC"].unique())) + " cells")
logger.info("Filtering out low UMI cell barcodes...")
allele_table = utilities.filter_cells(
allele_table,
min_umi_per_cell=int(min_umi_per_cell),
min_avg_reads_per_umi=min_avg_reads_per_umi,
)
allele_table["lineageGrp"] = allele_table["lineageGrp"].astype(int)
if plot:
logger.info("Producing Plots...")
at_pivot_I = pd.pivot_table(
allele_table,
index="cellBC",
columns="intBC",
values="UMI",
aggfunc="count",
)
at_pivot_I.fillna(value=0, inplace=True)
at_pivot_I[at_pivot_I > 0] = 1
logger.info("Producing pivot table heatmap...")
lineage_utils.plot_overlap_heatmap(
allele_table, at_pivot_I, output_directory
)
logger.info("Plotting filtered lineage group pivot table heatmap...")
lineage_utils.plot_overlap_heatmap_lg(
allele_table, at_pivot_I, output_directory
)
return allele_table