Feature request / Configuration for Parse Biosciences GeneSelect data (Tag: pN)

[jasoncrucify]

Dear Dev,

I am currently using scReadCounts (v1.3.1) to analyze single-cell data generated with the Parse Biosciences MEGA platform combined with GeneSelect (targeted enrichment).

Given the nature of GeneSelect data, which provides high coverage but remains sparse at the single-cell level, I need to use a minReads = 1 threshold to maintain sensitivity. However, I have encountered an issue with UMI deduplication: our BAM files store the UMI information in the pN tag, but the current -U option only supports a fixed list (STARsolo, CellRanger, UMI-tools) which defaults to UB or UR.

Could you advise on the best way to pass a custom UMI tag (like pN) to the command line? If this is not currently supported, would it be possible to allow a string input for the -U parameter or add a specific profile for recent Parse/STARsolo outputs?

Thank you for your help and for this great tool.

[jasoncrucify]

here is my setting

sys.argv = [
'scReadCounts.py',
'-r', '$BAM',
'-s', '$VCF',
'-b', '$BARCODES',
'-C', 'STARsolo',
'-U', 'STARsolo',
'-o', '$OUTPUT_FILE',
'-m', '1',
'-F'
]

we have CB:Z:83_19_04__s5 this is a the single cell ID
pN:Z:TAGGTTCCTT this is UMI from PARSE
also. pS:Z:CH0000 also this is the sample ID

for setting -U choosing the best methods, I am unable to select a specific column from the bam file we analyzed

[edwardsnj]

The group.ini file(s) define these barcode extraction rules and is user editable. See: https://horvathlab.github.io/NGS/SCExecute/docs/Barcodes.html Hope this helps! Once you have verified a working rule for GeneSelect cell barcodes and UMIs please send it to us to add to the packaged configuration files for others.

[jasoncrucify]

I spent some time on using the parse data following your recommendations, the results are a bit complicated to interpret: I modify the group.ini to read the barcode of each sc of merged bam files using GeneSelect. Two layers of observations: 1. We used several hundreds of probes to amplify different regions of interest (different genes), most of the probes work, but quite heterogeneously 2) We have a severe issue to count the goodreads of a single sample from merged BAM file: we first check using IGV to investigate the depth of coverage of a specific mutation of interest, and we try to target this specific mutation using

we try to align between the reads of IGV and the screadcount on several specific positions: however, some are aligned, but some of them not, and we seem to have a very low recovery rate of cell barcode through screadcount despite the results are quite expected and promising. I am wondering if there is a problem of running the script or setting? Please see attached one script.

THREADS=40

PATIENT_SAMPLE="XXXXXX"
SAMPLE_PREFIX="TO_FILL"
TARGET_GENE="ALB"
TARGET_PROT="p.I537T"
TARGET_CHROM="chr4"
TARGET_POS="73418269"
TARGET_REF="T"
TARGET_ALT="C"

MUT_ID="ALB_pI537T"

LOC="${TARGET_CHROM}:${TARGET_POS}-${TARGET_POS}"

OUT_DIR="/mnt/workdir/Workdir_ZZ/SCReadCounts/Analysis/Results/${MUT_ID}_all96_no_whitelist"
LOG_DIR="${OUT_DIR}/logs"
AUDIT_DIR="${OUT_DIR}/audit"

mkdir -p "$OUT_DIR" "$LOG_DIR" "$AUDIT_DIR"

TARGET_SNV="${OUT_DIR}/${MUT_ID}target_only.tsv"
SC_OUT="${OUT_DIR}/scReadCounts${MUT_ID}all96_no_whitelist_results.tsv"
SC_LOG="${LOG_DIR}/scReadCounts${MUT_ID}_all96_no_whitelist.log"

CB_PS_AUDIT="${AUDIT_DIR}/${MUT_ID}_CB_pS_locus_audit.tsv"
SUMMARY_OUT="${OUT_DIR}/${MUT_ID}_all96_sample_distribution_summary.tsv"

export OMP_NUM_THREADS=1
export OPENBLAS_NUM_THREADS=1
export MKL_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1

echo "=============================================================================="
echo "ALL96 no-whitelist scReadCounts"
echo "Mutation: ${TARGET_GENE} ${TARGET_PROT} ${TARGET_CHROM}:${TARGET_POS} ${TARGET_REF}>${TARGET_ALT}"
echo "BAM: $BAM"
echo "Output: $SC_OUT"
echo "=============================================================================="
date

==============================================================================

1. Sanity checks

==============================================================================

if [ ! -f "$PYTHON_BIN" ]; then
echo "ERROR: PYTHON_BIN not found: $PYTHON_BIN"
exit 1
fi

if [ ! -d "$WORKING_SRC_ORIGINAL" ]; then
echo "ERROR: original scReadCounts source folder not found: $WORKING_SRC_ORIGINAL"
exit 1
fi

if [ ! -f "$BAM" ]; then
echo "ERROR: BAM not found: $BAM"
exit 1
fi

if [ ! -f "${BAM}.bai" ]; then
echo "ERROR: BAM index not found: ${BAM}.bai"
exit 1
fi

if ! command -v samtools >/dev/null 2>&1; then
echo "ERROR: samtools not found in PATH."
exit 1
fi

==============================================================================

2. Create patched scReadCounts source copy

==============================================================================

echo "Creating patched scReadCounts source copy..."

rm -rf "$WORKING_ROOT_TEST"
mkdir -p "$WORKING_ROOT_TEST"
cp -a "$WORKING_SRC_ORIGINAL" "$WORKING_ROOT_TEST/"

PATCH_FILES=$(grep -RlE "skip_duplicate|skip_secondary|skip_qcfail|skip_unmapped" "$WORKING_SRC_TEST" || true)

if [ -n "$PATCH_FILES" ]; then
for f in $PATCH_FILES; do
perl -pi -e 's/(skip_duplicate\s*[=:]\s*)True/${1}False/g' "$f"
perl -pi -e 's/(skip_secondary\s*[=:]\s*)True/${1}False/g' "$f"
perl -pi -e 's/(skip_qcfail\s*[=:]\s*)True/${1}False/g' "$f"
perl -pi -e 's/(skip_unmapped\s*[=:]\s*)True/${1}False/g' "$f"
done
fi

==============================================================================

3. Build target-only SNV file

==============================================================================

{
echo -e "CHROM\tPOS\tREF\tALT"
echo -e "${TARGET_CHROM}\t${TARGET_POS}\t${TARGET_REF}\t${TARGET_ALT}"
} > "$TARGET_SNV"

cat "$TARGET_SNV"

==============================================================================

4. Locus audit: CB + pS mapping at this mutation locus

==============================================================================

echo "Building CB/pS locus audit..."

samtools view -@ "$THREADS" "$BAM" "$LOC"
| awk '
BEGIN {
FS = "\t";
OFS = "\t";
print "CB", "pS", "reads_at_locus";
}
{
cb = "NA";
ps = "NA";

for (i = 12; i <= NF; i++) {
    if ($i ~ /^CB:Z:/) {
        split($i, b, ":");
        cb = b[3];
    }
    if ($i ~ /^pS:Z:/) {
        split($i, p, ":");
        ps = p[3];
    }
}

if (cb != "NA") {
    key = cb SUBSEP ps;
    count[key]++;
}

}
END {
for (k in count) {
split(k, a, SUBSEP);
print a[1], a[2], count[k];
}
}
' > "$CB_PS_AUDIT"

echo "CB/pS audit preview:"
head "$CB_PS_AUDIT" | column -t || head "$CB_PS_AUDIT"

==============================================================================

5. Run scReadCounts WITHOUT whitelist

==============================================================================

echo "Running scReadCounts WITHOUT whitelist..."

$PYTHON_BIN -c "
import configparser, sys, runpy
configparser.SafeConfigParser = configparser.ConfigParser
sys.modules['configparser'].SafeConfigParser = configparser.ConfigParser
sys.path.insert(0, '$WORKING_SRC_TEST')

sys.argv = [
'scReadCounts.py',
'-r', '$BAM',
'-s', '$TARGET_SNV',
'-C', 'GeneSelect',
'-o', '$SC_OUT',
'-m', '0',
'-t', '$THREADS',
'-F'
]

runpy.run_path('$SC_PY', run_name='main')
" 2>&1 | tee "$SC_LOG"

echo "scReadCounts output:"
ls -lh "$SC_OUT"

==============================================================================

6. Summarize to all96 sample-level distribution using pS tag

==============================================================================

echo "Summarizing sample-level distribution..."

$PYTHON_BIN - <<PY
import csv
from collections import defaultdict, Counter
from pathlib import Path

sc_out = Path("$SC_OUT")
audit = Path("$CB_PS_AUDIT")
summary_out = Path("$SUMMARY_OUT")

target_chrom = "$TARGET_CHROM".replace("chr", "")
target_pos = int("$TARGET_POS")
target_ref = "$TARGET_REF"
target_alt = "$TARGET_ALT"

CB -> best pS at locus

cb_ps_counts = defaultdict(Counter)

with audit.open() as f:
reader = csv.DictReader(f, delimiter="\t")
for row in reader:
cb = row.get("CB", "")
ps = row.get("pS", "NA")
n = int(float(row.get("reads_at_locus", 0) or 0))
if cb and cb != "NA":
cb_ps_counts[cb][ps] += n

cb_to_ps = {}
for cb, ctr in cb_ps_counts.items():
cb_to_ps[cb] = ctr.most_common(1)[0][0]

sample_stats = defaultdict(lambda: {
"scReadCounts_REF_reads": 0,
"scReadCounts_ALT_reads": 0,
"scReadCounts_GoodReads": 0,
"scReadCounts_covered_CB": 0,
"scReadCounts_ALT_positive_CB": 0,
"scReadCounts_REF_positive_CB": 0,
"n_scReadCounts_rows": 0,
})

with sc_out.open() as f:
reader = csv.DictReader(f, delimiter="\t")
for row in reader:
chrom = str(row.get("CHROM", "")).replace("chr", "")
try:
pos = int(float(row.get("POS", "0")))
except Exception:
continue

    if chrom != target_chrom:
        continue
    if pos != target_pos:
        continue
    if str(row.get("REF", "")) != target_ref:
        continue
    if str(row.get("ALT", "")) != target_alt:
        continue

    cb = str(row.get("ReadGroup", "NA"))
    sample = cb_to_ps.get(cb, "UNKNOWN_pS")

    ref = int(float(row.get("RefCount", 0) or 0))
    alt = int(float(row.get("SNVCount", 0) or 0))
    good = int(float(row.get("GoodReads", 0) or 0))

    st = sample_stats[sample]
    st["scReadCounts_REF_reads"] += ref
    st["scReadCounts_ALT_reads"] += alt
    st["scReadCounts_GoodReads"] += good
    st["n_scReadCounts_rows"] += 1

    if good > 0:
        st["scReadCounts_covered_CB"] += 1
    if alt > 0:
        st["scReadCounts_ALT_positive_CB"] += 1
    if ref > 0:
        st["scReadCounts_REF_positive_CB"] += 1

rows = []
for sample, st in sample_stats.items():
good = st["scReadCounts_GoodReads"]
alt = st["scReadCounts_ALT_reads"]
ref = st["scReadCounts_REF_reads"]

rows.append({
    "sample_id": sample,
    "patient_sample": "$PATIENT_SAMPLE",
    "gene": "$TARGET_GENE",
    "protein": "$TARGET_PROT",
    "CHROM": "$TARGET_CHROM",
    "POS": "$TARGET_POS",
    "REF": "$TARGET_REF",
    "ALT": "$TARGET_ALT",
    "scReadCounts_REF_reads": ref,
    "scReadCounts_ALT_reads": alt,
    "scReadCounts_GoodReads": good,
    "scReadCounts_covered_CB": st["scReadCounts_covered_CB"],
    "scReadCounts_ALT_positive_CB": st["scReadCounts_ALT_positive_CB"],
    "scReadCounts_REF_positive_CB": st["scReadCounts_REF_positive_CB"],
    "n_scReadCounts_rows": st["n_scReadCounts_rows"],
    "ALT_fraction_reads": (alt / good) if good > 0 else 0
})

rows = sorted(rows, key=lambda x: (-x["scReadCounts_GoodReads"], -x["scReadCounts_ALT_reads"], x["sample_id"]))

fieldnames = [
"sample_id", "patient_sample", "gene", "protein", "CHROM", "POS", "REF", "ALT",
"scReadCounts_REF_reads", "scReadCounts_ALT_reads", "scReadCounts_GoodReads",
"scReadCounts_covered_CB", "scReadCounts_ALT_positive_CB",
"scReadCounts_REF_positive_CB", "n_scReadCounts_rows", "ALT_fraction_reads"
]

with summary_out.open("w", newline="") as f:
writer = csv.DictWriter(f, fieldnames=fieldnames, delimiter="\t")
writer.writeheader()
writer.writerows(rows)

print(f"Wrote sample-level summary: {summary_out}")
print(f"Number of pS/sample groups: {len(rows)}")
for r in rows[:10]:
print(r)
PY

echo "Done."
echo "Main outputs:"
echo "$SC_OUT"
echo "$CB_PS_AUDIT"
echo "$SUMMARY_OUT"
date

Correct me if there is some errors of coding?

Many thanks

[edwardsnj]

@jasoncrucify this copy and paste of your script isn't reproducing well in the issue thread. Please attach a file (script?) if you want me to read and understand it. Please attach the scReadCounts output (it contains lots of logging information) too. I can see you are modifying the source code to eliminate certain types of pathological cases being filtered out, which I don't recommend. You can use the -f option to change filtering, including using "None" to indicate no filtering be carried out. Also, the filter.ini file (similar to the groups.ini file, a local copy can add or override filter definitions) allows you to tweak filtering criteria. You have sample barcodes too, which we don't have any specific way to handle - perhaps these could be rolled into the cell-barcodes in the group definition, but this might need a code update, or maybe it would be possible to define a read-filter that would remove anything that isn't from a specific sample. Lastly, scReadCounts is just a wrapper around the readCounts tool (the command is mentioned in the logging output) - and the readCounts tool has an option "-E' for extended output. One option "Read filtering statistics" can tell you how many reads were removed / filtered out for what reason. Perhaps this would offer some insight.

[jasoncrucify]

Thanks Yes indeed, I was testing the 4 QC filters you applied. I think we introduce quite a lot of duplicates. Indeed, I am testing several parameters in the arguments in particular -C barcodes, -U, UMI (set for gene select, giving a specific tag from bam file) modified in group.ini. I think removing the 4 filters seem to have false positive results (some not-supposed-to-be mutated cells are mutated. Regarding the minimum read (-m), how to define the threshold? Any criteria to apply? I tried 0 or 1, both outputs the same number of cells (identical barcode number) but less goodreadcounts (REF and ALT)? I am wondering what is the threshold considered as true positive? We expect for some mutations (i.e. 0.3-0.5 VAFs), we should not see all the cells are mutated, but now it is the case. Many thanks for your reply.

…

Le 26 juin 2026 à 19:22, Nathan Edwards ***@***.***> a écrit : edwardsnj left a comment (HorvathLab/NGS#70) <#70 (comment)> @jasoncrucify <https://github.com/jasoncrucify> this copy and paste of your script isn't reproducing well in the issue thread. Please attach a file (script?) if you want me to read and understand it. Please attach the scReadCounts output (it contains lots of logging information) too. I can see you are modifying the source code to eliminate certain types of pathological cases being filtered out, which I don't recommend. You can use the -f option to change filtering, including using "None" to indicate no filtering be carried out. Also, the filter.ini file (similar to the groups.ini file, a local copy can add or override filter definitions) allows you to tweak filtering criteria. You have sample barcodes too, which we don't have any specific way to handle - perhaps these could be rolled into the cell-barcodes in the group definition, but this might need a code update, or maybe it would be possible to define a read-filter that would remove anything that isn't from a specific sample. Lastly, scReadCounts is just a wrapper around the readCounts tool (the command is mentioned in the logging output) - and the readCounts tool has an option "-E' for extended output. One option "Read filtering statistics" can tell you how many reads were removed / filtered out for what reason. Perhaps this would offer some insight. — Reply to this email directly, view it on GitHub <#70?email_source=notifications&email_token=ANKIWF6BYJDDSKBAZ6F5C435B2WLZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBRGE4DAOJZGE4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4811809918>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANKIWF4GBJ4ZZW2JIC2BRJD5B2WLZAVCNFSNUABEKJSXA33TNF2G64TZHM2DINZVGQYTMMR3JFZXG5LFHM2DIMBSGE2DIMRVGOQXMAQ>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ANKIWF5ZDYF4XX4RZT3GPYD5B2WLZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBRGE4DAOJZGE4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/ANKIWFYYEAJ24SL33Z5RFW35B2WLZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBRGE4DAOJZGE4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.

[jasoncrucify]

Final_ALB_pI537T_no_whitelist_stringent_filte.sh here is the script

[jasoncrucify]

Dear Dev

I hope you are well.

I am currently using scReadCounts to quantify a specific target SNV in single-cell / single-nucleus sequencing BAM files generated from Parse Biosciences data, and I would like to ask for your advice on an apparent discrepancy between raw BAM/IGV-level read depth and scReadCounts GoodReads / RefCount / SNVCount.

The target variant is:

Gene: ALB
Protein: p.I537T
Locus: chr4:73418269 T>C
REF: T
ALT: C
Sample: CHC1600T

I am running scReadCounts on a per-sample split BAM:

parse_CHC1600T.bam

with a one-row SNV file:

CHROM POS REF ALT
chr4 73418269 T C

The basic scReadCounts command structure is:

sys.argv = [
"scReadCounts.py",
"-r", BAM,
"-s", TARGET_SNV,
"-C", "GeneSelect",
"-o", SC_OUT,
"-m", "0",
"-t", THREADS,
"-F"
]

I also created a patched test version of the source where the four basic read-skip filters were disabled:

skip_duplicate = False
skip_secondary = False
skip_qcfail = False
skip_unmapped = False

The goal was to test whether the lower scReadCounts counts compared with IGV/raw BAM were due to duplicate, secondary, QC-fail, or unmapped read filtering.

The main result is that disabling these four filters had almost no effect. For CHC1600T, the no-basic-filter scReadCounts output was:

scReadCounts RefCount = 66,305
scReadCounts SNVCount = 31,405
scReadCounts GoodReads = 97,754
covered CBs = 8,751
ALT-positive CBs = 2,574

However, an independent BAM audit at the same locus gave:

raw overlapping reads with CB/pS tags = 114,505
aligned-base reads at target = 112,243
deletion/refskip reads = 2,262
raw REF / T reads = 76,763
raw ALT / C reads = 35,428
other base reads = 52
raw REF+ALT informative reads = 112,191

So even after excluding deletion/refskip reads, other-base reads, and non-informative reads, the raw BAM still contains:

112,191 REF/ALT-informative reads

whereas scReadCounts reports:

66,305 + 31,405 = 97,710 REF+ALT reads

This leaves approximately:

112,191 - 97,710 = 14,481 REF/ALT-informative reads

that are present in the raw BAM audit but not counted in scReadCounts RefCount or SNVCount.

I also checked barcode/sample tags. The raw overlapping reads all had usable tags:

CB present: 114,505 / 114,505
CR present: 114,505 / 114,505
pS present: 114,505 / 114,505
pS value: CHC1600T
RG absent: all reads
UB absent: all reads

Therefore, the discrepancy does not appear to be explained by missing cell barcodes or sample tags.

My current interpretation is that scReadCounts is applying additional GoodReads / allele-informative read logic beyond the four basic skip filters. Possible explanations I am considering include:

1. paired-end mate overlap handling, where raw BAM/IGV counts both mates but scReadCounts avoids double-counting one fragment;
2. molecule-level or UMI-like collapsing, possibly related to the -U argument if used;
3. base-quality or mapping-quality thresholds internal to scReadCounts;
4. CIGAR/read-position handling near the target base;
5. another internal GoodReads criterion that I have not identified.

One specific question: in some tests, I used:

"-U", "GeneSelect"

I am unsure whether -U GeneSelect triggers UMI/molecule-level collapsing or another form of deduplication. Since this BAM does not appear to contain standard UB tags, I would like to clarify what -U does in this context.

Could you please help clarify the following?

1. What is the exact definition of GoodReads in scReadCounts?
2. Does scReadCounts count raw alignments, read pairs/fragments, or UMI/molecule-collapsed observations?
3. Does -U GeneSelect collapse reads that share the same molecule/UMI-like identity?
4. Are there internal base-quality, mapping-quality, read-end, or CIGAR filters applied before reads enter GoodReads?
5. In a high-depth target site like this, would it be expected that raw BAM REF/ALT-informative reads exceed scReadCounts RefCount + SNVCount by ~13%?
6. Is there a recommended diagnostic mode or output to inspect reads rejected before GoodReads assignment?

For reproducibility, the main scripts I used are:

1. Per-sample scReadCounts batch script with original/default filtering.
2. Patched no-basic-filter version where:
   skip_duplicate = False
   skip_secondary = False
   skip_qcfail = False
   skip_unmapped = False
3. BAM read-fate audit script that classifies each raw overlapping read by:
   CB/pS tag status,
   read flags,
   CIGAR/target-base status,
   allele class at chr4:73418269,
   REF/ALT/other/deletion/refskip category.
4. Per-CB reconciliation script comparing:
   raw BAM REF/ALT counts per CB
   versus scReadCounts RefCount/SNVCount per CB.

The key comparison is:

raw BAM REF+ALT-informative reads: 112,191
scReadCounts RefCount + SNVCount: 97,710
missing REF/ALT reads: ~14,481

I would appreciate your guidance on which internal scReadCounts rule most likely explains this gap, and whether my interpretation of GoodReads is correct.

Many thanks for your help.

[edwardsnj]

I think removing the 4 filters seem to have false positive results (some not-supposed-to-be mutated cells are mutated.

Yes, these basic filters are always applied because reads that align with these flags are probably not good evidence for the sequence at that locus (the correctness of the alignment of the read to that spot is in doubt), so any variants observed may just be because the sequence is misaligned.

Regarding the minimum read (-m), how to define the threshold? Any criteria to apply? I tried 0 or 1, both outputs the same number of cells (identical barcode number) but less goodreadcounts (REF and ALT)? I am wondering what is the threshold considered as true positive?

There is no threshold that should be considered a true positive, but certainly 0 or 1 is unlikely to be good (essentially no filter).

We expect for some mutations (i.e. 0.3-0.5 VAFs), we should not see all the cells are mutated, but now it is the case.

Yes, you are setting options that allow a huge amount of noise to be observed.

[edwardsnj]

Lastly, scReadCounts is just a wrapper around the readCounts tool (the command is mentioned in the logging output) - and the readCounts tool has an option "-E' for extended output. One option "Read filtering statistics" can tell you how many reads were removed / filtered out for what reason. Perhaps this would offer some insight.

Could you please do this experiment for me on your small example? Check the logging output, find the readCounts command, and run it with the recommended -E option. This provides a complete accounting of which reads are considered Good and if not, what rule decided to discard them. Alternatively, if you can provide me with a small section of your BAM file that demonstrates the issue, I will attempt to reproduce your issue.

[edwardsnj]

There is so much here, it is hard to cut to the essence of the issue. However, I have spent some time looking over the "Read filtering statistics" for the -E extended output option for readCounts (which is the source of any / all read counts used by scReadCounts).

Here is the essential issue, I think. You are using readCounts/scReadCounts to count # distinct UMIs (-U option) NOT number of reads. This is good, since UMIs help remove the noise due to duplicate read artifacts.

Your various methods for checking, including the one that looks at the reads directly, do not account for counting distinct UMIs.

However, I don't think, ultimately this is the core of the issue that sent you down this rabbit-hole, since the VAR from your analysis and scReadCounts is about the same ~ 0.321 vs 0.316. I think the core of the issue is that you think you have too many false negatives. If your data is really this sparse enough, on a per cell level, to require min reads == 1, then many sites will just go unobserved (no reads at all for a specific cell), or have VAR 0 or 1 (necessarily). Note too, that UMI-based counting won't drop below 1 if any read is observed - since that read has a UMI.

Hope all of this helps...