GPTomics/bio-duplicate-handling

Version Compatibility

Reference examples tested with: picard 3.1+, pysam 0.22+, samtools 1.19+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Duplicate Handling

"Remove PCR duplicates from my BAM file" -> Mark or remove duplicate reads using the fixmate-sort-markdup pipeline to prevent duplicate bias in variant calling.

• CLI: samtools fixmate, samtools markdup (samtools)
• Python: pysam.fixmate(), pysam.markdup() (pysam)

Mark and remove PCR/optical duplicates using samtools.

Why Remove Duplicates?

PCR duplicates are identical copies of the same original molecule, created during library preparation. They inflate coverage, bias allele frequencies, and create false positive variant calls. Optical duplicates are flowcell-proximity artifacts: on unpatterned flowcells they arise when the imaging software splits one cluster into two adjacent calls; on patterned flowcells (NovaSeq, NovaSeq X, NextSeq 1000/2000, HiSeq X/4000) the dominant source is ExAmp (exclusion-amplification) "pad-hopping", where a library molecule re-seeds a nearby nanowell.

When to Mark Duplicates -- and When NOT To

Standard samtools markdup is the right tool for some assays and actively harmful for others. The decision is assay-driven:

| Assay | Standard markdup? | Recommended approach | |-------|------------------|----------------------| | Germline WGS / WES (PCR or PCR-free) | YES | samtools markdup (PCR-free still has ~0.5% optical duplicates on patterned flowcells) | | Somatic tumor/normal (no UMI) | YES | Same | | Exome / target capture | YES (20-50% expected) | samtools markdup | | ChIP-seq | MARK, do not remove | Then use peak caller's auto-dup logic (macs3 --keep-dup auto) | | CUT&RUN / CUT&Tag | MARK, do not remove | Same | | ATAC-seq | YES, BEFORE Tn5 +4/-5 shift | Then shift coords for footprinting | | Bulk RNA-seq (no UMIs) | NO | Duplicates are biological at highly-expressed loci; removing them biases DE proportional to expression | | Bulk RNA-seq (with UMIs) | NO | umi_tools dedup | | scRNA (10x, STARsolo, drop-seq) | NO | umi_tools dedup with CB+UB tags, or rely on Cell Ranger UMI counts | | ctDNA / liquid biopsy / deep panel (UMI) | NO | fgbio GroupReadsByUmi -> CallDuplexConsensusReads | | Twist / IDT / Roche UMI capture | NO | fgbio or Picard UmiAwareMarkDuplicatesWithMateCigar | | Amplicon / hotspot panel (no UMI) | NO | Every read is a "duplicate" by coordinate; markdup erases the dataset. Use samtools ampliconclip instead -- see alignment-amplicon-clipping. | | Amplicon / hotspot panel (UMI) | NO | fgbio consensus | | Long-read native (ONT, PacBio HiFi unamplified) | NO | No PCR step; markdup is meaningless | | PacBio HiFi amplicon | YES (rare) | pbmarkdup | | Ancient DNA (aDNA) | YES + mapDamage | Run markdup, then mapDamage --rescale before variant calling | | Microbiome 16S/ITS | NO | Read counts encode community structure |

If the BAM came from 10x Cell Ranger / STARsolo and samtools markdup produces a 50-95% duplicate rate, it is the wrong tool, not a bug.

Tool Selection: markdup vs Picard vs UMI-aware

| Tool | Speed | Threading | Optical | UMI | Notes | |------|-------|-----------|---------|-----|-------| | samtools markdup | Fast | Yes | Yes (-d) | Limited (--barcode-tag exact-match) | Fast production choice (nf-core/sarek defaults to GATK MarkDuplicates) | | picard MarkDuplicates | Slow | No | Yes | UmiAware variant (BETA, transcriptome bug) | GATK Best Practices reference | | biobambam2 bammarkduplicates2 | Fastest | Yes | Yes | No | Sanger / 1KGP pipelines | | samblaster | Streaming, fast | No | Optional | No | Pipe directly from aligner; no name sort | | sambamba markdup | Fast | Yes | Yes | No | Less actively maintained | | fgbio GroupReadsByUmi + CallMolecularConsensusReads | Fast | Yes | n/a | Best UMI tool | Graph-based; supports duplex | | umi_tools dedup | Slow | No | n/a | Yes (mature) | Reference for scRNA / bulk UMI | | pbmarkdup | Fast | Yes | n/a | n/a | PacBio HiFi amplicons only |

Picard UmiAwareMarkDuplicatesWithMateCigar is BETA and has known bugs on transcriptome-aligned BAMs (silently keeps duplicates). Avoid for RNA-seq UMIs.

Optical Distance Is Platform-Specific

samtools markdup default is -d 0, meaning optical-duplicate detection is disabled by default. Set explicitly per platform:

| Platform | -d value | Rationale | |----------|-----------|-----------| | HiSeq 2000/2500 (random) | 100 | Picard historic default | | HiSeq 3000/4000/X (patterned) | 2500 | Patterned tile size larger | | NovaSeq 6000 (patterned) | 2500 | Same as HiSeq X | | NovaSeq X (10B) | 2500 | Patterned; same starting point as NovaSeq 6000 | | NextSeq 1000/2000 (patterned) | 2500 | ExAmp duplicates span larger pixel distances | | MiSeq, NextSeq 500/550 | 100 | Smaller / unpatterned | | Element AVITI, MGI / DNBseq | Custom regex | Different read-name format -- supply via --read-coords |

samtools markdup -d 2500 -f stats.txt input.bam marked.bam

# Count optical (SQ) vs library/PCR (LB) duplicates. The dt:Z:SQ/LB tag is
# emitted automatically because -d is set (it is not produced by -t, which
# instead adds a 'do' tag carrying the original read's name).
samtools view -f 1024 marked.bam | grep -o 'dt:Z:[A-Z][A-Z]' | sort | uniq -c

Setting -d 2500 on a HiSeq run does no harm. Forgetting -d 2500 on NovaSeq systematically under-marks optical duplicates and overestimates library complexity.

Multi-Library Pooled Marking

Without --use-read-groups, multi-library BAMs systematically over-mark: independent molecules from different libraries with the same coordinates get wrongly flagged as PCR duplicates. With --use-read-groups, RG tags must also match for two reads to be a duplicate (verify availability with samtools markdup --help):

samtools markdup --use-read-groups -d 2500 in.bam out.bam

samtools --use-read-groups keys on RG ID; Picard's library-aware behavior keys on the LB tag (allowing dedup across multiple lanes of the same library). For multi-lane single-library BAMs, Picard MarkDuplicates with READ_NAME_REGEX is closer to canonical.

Duplicate Marking Workflow

Goal: Mark PCR/optical duplicates so they can be excluded from downstream variant calling and coverage analysis.

Approach: Name-sort, add mate tags with fixmate, coordinate-sort, then run markdup. The pipeline version avoids intermediate files.

Reference (samtools 1.19+):

# 1. Sort by name (required for fixmate)
samtools sort -n -o namesort.bam input.bam

# 2. Add mate information with fixmate
samtools fixmate -m namesort.bam fixmate.bam

# 3. Sort by coordinate (required for markdup)
samtools sort -o coordsort.bam fixmate.bam

# 4. Mark duplicates
samtools markdup coordsort.bam marked.bam

# 5. Index result
samtools index marked.bam

Pipeline Version (Optimized)

# collate is faster than sort -n; -u/-O between piped tools skips BGZF round-trips
samtools collate -O -u input.bam tmpdir/collate | \
    samtools fixmate -m -u - - | \
    samtools sort -u -@ 4 -T tmpdir/sort - | \
    samtools markdup -@ 4 -d 2500 --use-read-groups \
        -f markdup_stats.txt - marked.bam

samtools index marked.bam

This is ~30% faster than sort -n | fixmate | sort | markdup on typical 30x WGS.

Critical pitfall: samtools markdup requires ms (mate score, lowercase) and MC (mate CIGAR) tags from fixmate -m. A re-sort that loses aux tags via Python round-trip silently produces a markdup output that marks almost nothing. If duplicate counts look implausibly low, verify MC:Z: is present in the input to markdup.

samtools fixmate

Adds mate information required by markdup. Must be run on name-sorted BAM.

Basic Usage

samtools fixmate namesorted.bam fixmate.bam

Add Mate Score Tag (-m)

# Required for markdup to work correctly
samtools fixmate -m namesorted.bam fixmate.bam

Multi-threaded

samtools fixmate -m -@ 4 namesorted.bam fixmate.bam

Remove Secondary/Unmapped

samtools fixmate -r -m namesorted.bam fixmate.bam

samtools markdup

Marks or removes duplicate alignments. Requires coordinate-sorted BAM with mate tags from fixmate.

Mark Duplicates (Keep in File)

samtools markdup input.bam marked.bam

Remove Duplicates

samtools markdup -r input.bam deduped.bam

Output Statistics

samtools markdup -s input.bam marked.bam 2> markdup_stats.txt

Optical Duplicate Distance

# Default -d 0 disables optical detection. Set per platform; see decision table above.
samtools markdup -d 2500 input.bam marked.bam   # NovaSeq / patterned
samtools markdup -d 100 input.bam marked.bam    # HiSeq / random

Multi-threaded

samtools markdup -@ 4 input.bam marked.bam

Write Stats to File

samtools markdup -f stats.txt input.bam marked.bam

Duplicate Statistics

Check Duplicate Rate

samtools flagstat marked.bam
# Look for "duplicates" line

Count Duplicates

# Count reads with duplicate flag
samtools view -c -f 1024 marked.bam

Percentage Duplicates

total=$(samtools view -c marked.bam)
dups=$(samtools view -c -f 1024 marked.bam)
echo "scale=2; $dups * 100 / $total" | bc

pysam Python Alternative

Full Pipeline

import pysam

# Sort by name
pysam.sort('-n', '-o', 'namesort.bam', 'input.bam')

# Fixmate
pysam.fixmate('-m', 'namesort.bam', 'fixmate.bam')

# Sort by coordinate
pysam.sort('-o', 'coordsort.bam', 'fixmate.bam')

# Mark duplicates
pysam.markdup('coordsort.bam', 'marked.bam')

# Index
pysam.index('marked.bam')

Check Duplicate Flag

import pysam

with pysam.AlignmentFile('marked.bam', 'rb') as bam:
    total = 0
    duplicates = 0
    for read in bam:
        total += 1
        if read.is_duplicate:
            duplicates += 1

    print(f'Total: {total}')
    print(f'Duplicates: {duplicates}')
    print(f'Rate: {duplicates/total*100:.2f}%')

Filter Out Duplicates

import pysam

with pysam.AlignmentFile('marked.bam', 'rb') as infile:
    with pysam.AlignmentFile('nodup.bam', 'wb', header=infile.header) as outfile:
        for read in infile:
            if not read.is_duplicate:
                outfile.write(read)

Production Tools, Not Hand-Rolled

For real BAMs, always use a production marker. A naive Python implementation keyed on (chrom, pos, strand) ignores 5' position correction for soft clips, ignores library/RG, treats optical = PCR, and mis-handles supplementary alignments. The result is silently wrong duplicate marks. Use samtools markdup, Picard, or fgbio depending on assay (see decision tables above).

Alternative: From Aligner

Some aligners can mark duplicates directly during streaming:

BWA-MEM2 with samblaster

bwa-mem2 mem ref.fa R1.fq R2.fq | \
    samblaster | \
    samtools sort -o marked.bam

Picard MarkDuplicates

java -jar picard.jar MarkDuplicates \
    I=input.bam \
    O=marked.bam \
    M=metrics.txt \
    OPTICAL_DUPLICATE_PIXEL_DISTANCE=2500

UMI-Aware Deduplication

For UMI libraries (10x scRNA, ctDNA panels, Twist/IDT/Roche UMI capture), naive markdup destroys information. Use UMI-aware tools:

# 10x / scRNA -- group by cell barcode + UMI
umi_tools dedup --stdin=cellranger_possorted.bam --stdout=dedup.bam \
    --extract-umi-method=tag --umi-tag=UB --cell-tag=CB \
    --per-cell --method=directional

# Bulk UMI / ctDNA -- consensus calling (best practice for low-VAF detection)
fgbio AnnotateBamWithUmis -i raw.bam -f umi.fastq -o annotated.bam
fgbio GroupReadsByUmi -i annotated.bam -o grouped.bam --strategy=adjacency --edits=1
fgbio CallMolecularConsensusReads -i grouped.bam -o consensus.bam --min-reads=1
# Or for duplex (xGen-Prism, NEBNext duplex):
fgbio CallDuplexConsensusReads -i grouped.bam -o duplex.bam --min-reads 1 1 0

--method=directional is the default and correct -- do not use --method=unique, which treats single-base UMI errors as different molecules. samtools markdup --barcode-tag RX (UMI/barcode handling added in samtools 1.16) does exact-match UMI grouping; adequate for IDT xGen Duplex but insufficient for single-UMI applications where 1-edit errors are common.

Quick Reference

| Task | Command | |------|---------| | Full workflow | sort -n \| fixmate -m \| sort \| markdup | | Mark duplicates | samtools markdup in.bam out.bam | | Remove duplicates | samtools markdup -r in.bam out.bam | | Count duplicates | samtools view -c -f 1024 marked.bam | | View non-duplicates | samtools view -F 1024 marked.bam | | Get stats | samtools markdup -s in.bam out.bam |

Duplicate FLAG

| Flag | Value | Meaning | |------|-------|---------| | 0x400 | 1024 | PCR or optical duplicate |

Filter Commands

# View only duplicates
samtools view -f 1024 marked.bam

# View non-duplicates only
samtools view -F 1024 marked.bam

# Count non-duplicates
samtools view -c -F 1024 marked.bam

Common Errors

| Error | Cause | Solution | |-------|-------|----------| | mate not found | Input not name-sorted | Run samtools sort -n first | | no MC tag | fixmate not run with -m | Re-run fixmate with -m flag | | not coordinate sorted | Input to markdup not sorted | Run samtools sort after fixmate |

Lossy Operations

samtools markdup -r (remove duplicates) is irreversible -- the records are dropped. Default to marking, not removing; downstream tools can filter on FLAG 1024. Removing pre-emptively destroys data needed for re-running QC, library complexity estimation, or switching dedup strategies.

Related Skills

• alignment-sorting - Sort by name/coordinate; collate vs sort -n decision
• alignment-filtering - Filter duplicates from output
• alignment-amplicon-clipping - Use ampliconclip instead of markdup for amplicon panels
• bam-statistics - Check duplicate rates with flagstat / mosdepth
• variant-calling/variant-calling - Standard variant calling expects deduped BAMs
• read-qc/quality-reports - Pre-alignment QC including UMI extraction