Contamination screening

Version Compatibility

Reference examples tested with: FastQ Screen 0.15+, Bowtie2 2.5+, Kraken2 2.1+, BBTools 39.0+, MultiQC 1.21+

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

• 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.

Contamination Screening -- a species screen cannot see a same-species swap

Screen reads against a genome panel (FastQ Screen / Kraken2) for foreign ORGANISMS, and against SNP fingerprints for foreign or wrong INDIVIDUALS.

"Check my reads for contamination" -> Map a subsample against multiple references and/or fingerprint sample identity to find foreign DNA and mislabels.

• CLI: fastq_screen --conf fastq_screen.conf sample.fastq.gz (cross-species)
• CLI: verifyBamID2 --SVDPrefix resource --BamFile sample.bam (same-species contamination)

Scope: this skill OWNS contamination detection and the report-vs-filter decision. Deep taxonomic profiling/abundance -> metagenomics/kraken-classification, metagenomics/metaphlan-profiling. rRNA depletion as an RNA prep metric -> read-qc/rnaseq-qc. OUT OF SCOPE: adapter removal (read-qc/adapter-trimming).

The Single Most Important Modern Insight

1. A SPECIES screen answers "what ORGANISMS are here?"; a SNP FINGERPRINT answers "WHOSE DNA is this, and is it a mixture?" -- and these are orthogonal. A species screen is structurally BLIND to same-species cross-sample contamination and sample swaps: a human-A + human-B mixture, or a mislabeled human file, produces a perfectly clean single-species profile. Most pipelines run only a species screen and declare the data clean. Any human or single-species-cohort pipeline needs BOTH a taxonomic screen AND a SNP-fingerprint identity/contamination check (verifyBamID2, NGSCheckMate, somalier, conpair).

2. Index hopping is the same-species contamination that lives inside one run, and unique dual indexing (UDI) is the only clean fix. On patterned flowcells (HiSeq X/4000, NovaSeq) ExAmp chemistry lets a free index adapter tag a fragment from another sample, spreading ~0.1-2% of reads into incorrect samples. Irrelevant for germline common variants, CATASTROPHIC for low-VAF work (ctDNA, single-cell, somatic) where hopped reads look like phantom low-frequency variants. Combinatorial indexing cannot detect it; UDI (a unique i7 AND i5 per sample) lets the demultiplexer drop impossible index pairs. UDI is effectively mandatory for ctDNA/plasma, single-cell, and low-input somatic.

3. Default to SCREEN-AND-REPORT, not filter -- removing reads biases composition. A species screen is a QC gate; if a contaminant is low, aligning to the correct reference simply will not place the foreign reads. Filter only a specific, named contaminant (PhiX before assembly, adapters before alignment) with a precise k-mer remover (BBDuk ref=phix), never "remove anything that hits the screen" (that also discards conserved rRNA/mito reads that belong to the sample). For PDX, align to a COMBINED human+mouse reference and keep human-assigned reads, OR use a dedicated post-alignment classifier (XenofilteR benchmarks above Xenome for variant false-positive rate); both beat hard pre-filtering on one genome, which mis-assigns conserved-region reads.

Deeper trap: reference-genome contamination corrupts the screen itself. A "human" hit can be bacterial sequence mis-deposited inside the human assembly (Conterminator found >2M contaminated GenBank entries). No --confidence setting fixes a wrong database; trust a deconned/curated DB and treat surprising single-source hits as DB artifacts until ruled out.

The Contamination Taxonomy -- five classes, five fixes

Class	What it is	Detect with	Fix
Cross-species	Mouse in human PDX; bacteria in culture	FastQ Screen, Kraken2/Bracken, sourmash, Xenome/XenofilteR	Combined-reference alignment; k-mer bin (report, do not blindly remove)
Cross-sample / index hopping	Same-species reads on the wrong sample (INVISIBLE to species screens)	verifyBamID2, NGSCheckMate, somalier, conpair	UDI at prep; drop impossible index pairs at demux
Vector / PhiX / adapter	Spike-in, cloning vector, linkers	UniVec/VecScreen; FastQ Screen adapter DB; BBDuk ref=phix/ref=adapters	k-mer trim/filter; upstream library QC
rRNA over-representation	Library-prep failure, not contamination	SortMeRNA, ribodetector	Re-prep / better depletion; filter only to recover depth
Cell-line: mycoplasma / misID	Mollicutes infection; HeLa cross-contamination	Kraken2 for Mollicutes; STR profiling for line identity	Clear culture or discard; STR-authenticate

A pipeline that runs FastQ Screen and stops has checked exactly one of five boxes.

Tool Taxonomy

Tool	Mechanism	When
FastQ Screen	Map a subsample to a genome panel; classify hit categories	Cross-species QC gate; the workhorse screen
Kraken2 + Bracken	Exact-k-mer minimizer LCA classification; Bracken re-estimates abundance	Read-level taxonomy; many possible contaminants
BBSplit / BBDuk	k-mer binning / named-contaminant k-mer removal	Decontamination of a NAMED contaminant (PhiX, adapters)
Xenome / XenofilteR	Classify reads human vs mouse (k-mer / dual-alignment)	PDX host-graft disambiguation
sourmash	MinHash/FracMinHash containment sketches	Fast low-memory "what is in here?" screen
verifyBamID2	Per-sample within-species contamination from population SNP AFs	Same-species contamination level (FREEMIX)
NGSCheckMate / somalier	SNP-fingerprint identity / relatedness	Sample swaps, tumor-normal pairing, longitudinal identity
conpair	Tumor-normal concordance + independent contamination	Matched T/N pairs

Decision Tree by Scenario

Question	Use	Why
Is a foreign ORGANISM present?	FastQ Screen or Kraken2	Maps reads to species references
Is this the right INDIVIDUAL / one person?	NGSCheckMate / somalier	SNP fingerprint, species-screen-blind
What is the contamination LEVEL (human)?	verifyBamID2 (FREEMIX)	Estimates mixture fraction from SNP AFs
Tumor-normal pair: matched and clean?	conpair	Concordance + per-sample contamination
PDX host vs graft	Combined reference or a benchmarked classifier	XenofilteR > Xenome for SNV FP rate; both beat hard pre-filtering
Remove a NAMED contaminant	BBDuk ref=...	Precise k-mer removal, not "hits the screen"
Strip HUMAN reads before public deposition (non-human library)	hostile / NCBI sra-human-scrubber (HRRT)	Deposition compliance, not a QC gate; a masked T2T reference avoids stripping conserved microbial regions

Default when uncertain: FastQ Screen as the QC gate for organisms, PLUS a SNP-fingerprint check (somalier/NGSCheckMate) for any human cohort.

FastQ Screen

Maps a SUBSAMPLE (--subset, default 100000) with bowtie2 reporting >1 alignment, then classifies each read across the panel. Read the bar chart, not just "% mapped": contamination concentrates in One_hit_one_genome of an UNEXPECTED genome; homology (rRNA, mito, conserved loci) spreads into the *_multiple_genomes categories; high Hit_no_genomes means adapter dimer, a missing reference, or a novel organism (a diagnostic, not a verdict).

# Config: aligner binary + DATABASE lines (bowtie2 index prefixes)
cat > fastq_screen.conf <<'EOF'
BOWTIE2  /usr/local/bin/bowtie2
THREADS  8
DATABASE  Human  /refs/GRCh38_bt2/GRCh38
DATABASE  Mouse  /refs/GRCm39_bt2/GRCm39
DATABASE  Ecoli  /refs/Ecoli_bt2/Ecoli
DATABASE  PhiX   /refs/phix_bt2/phix
DATABASE  rRNA   /refs/rRNA_bt2/rRNA
EOF

fastq_screen --conf fastq_screen.conf --threads 8 --outdir screen/ *.fastq.gz
multiqc screen/                      # MultiQC parses *_screen.txt across samples

# Tag every read with a per-genome status, then extract a subset by pattern
fastq_screen --conf fastq_screen.conf --tag --filter 10000 sample.fastq.gz   # maps only to genome 1
fastq_screen --conf fastq_screen.conf --nohits sample.fastq.gz               # reads hitting nothing

--filter digits (one per genome, config order): 0=no map, 1=unique, 2=multi, 3=maps, 4=pass 0 or 1, 5=pass 0 or 2, -=ignore. --subset 0 screens the whole file; --bisulfite uses Bismark.

Kraken2 + Bracken (read-level taxonomy)

Default --confidence 0.0 over-reports a long tail of spurious low-abundance species (a few shared k-mers suffice); raise to 0.05-0.1 and keep --minimum-hit-groups 2 (or 3 for custom DBs). Kraken2 gives CLASSIFICATION; Bracken redistributes higher-rank reads to species for ABUNDANCE.

kraken2 --db /db/k2_standard --threads 8 --confidence 0.1 --paired \
        --report sample.kreport --use-names R1.fq.gz R2.fq.gz > sample.kraken
bracken -d /db/k2_standard -i sample.kreport -o sample.bracken -r 150 -l S

Same-species: SNP fingerprints and index hopping

# verifyBamID2: FREEMIX = contamination fraction (action threshold ~0.02);
# FREEMIX~0 with CHIPMIX~1 indicates a SWAP, not contamination.
# --SVDPrefix points to the panel resource that ships with verifyBamID2 (resource/1000g.phase3...).
verifyBamID2 --SVDPrefix /res/1000g.phase3.100k.b38.vcf.gz.dat --BamFile sample.bam --Reference ref.fa

# somalier: extract genome sketches, then relate to find swaps / identity across a cohort.
# --sites = somalier's released sites.<build>.vcf.gz (github releases), not a custom panel.
somalier extract -d sites/ --sites sites.hg38.vcf.gz -f ref.fa sample.bam
somalier relate sites/*.somalier            # off-diagonal identity flags swaps

# conpair (tumor-normal): concordance + independent per-sample contamination

Index hopping is mitigated at demultiplexing with UDI (drop impossible i7,i5 pairs); residual contamination is then quantified by the SNP-fingerprint tools above.

Common Errors

Symptom	Cause	Solution
"Single species, data is clean" but a swap is suspected	Species screen is blind to same-species swaps	Run somalier / NGSCheckMate on SNP fingerprints
Phantom low-VAF variants in ctDNA/single-cell	Index hopping on patterned flowcell	Use UDI; quantify residual with verifyBamID2/conpair
Kraken2 reports dozens of trace species	Default confidence 0.0 over-reports	Raise --confidence to 0.05-0.1; raise hit-groups
A "human" Kraken hit on a microbial isolate	Reference/DB contamination	Use a deconned DB; treat as artifact until confirmed
Filtering "contaminant" reads skews composition	Removed conserved rRNA/mito too	Remove a NAMED contaminant with BBDuk, not "hits the screen"
PDX human counts look biased	Hard pre-filtering of ambiguous reads	Align to combined human+mouse reference instead

References

Wingett SW, Andrews S. 2018. FastQ Screen: a tool for multi-genome mapping and quality control. F1000Research 7:1338. Wood DE, Lu J, Langmead B. 2019. Improved metagenomic analysis with Kraken 2. Genome Biology 20:257. Lu J, Breitwieser FP, Thielen P, Salzberg SL. 2017. Bracken: estimating species abundance in metagenomics data. PeerJ Computer Science 3:e104. Steinegger M, Salzberg SL. 2020. Terminating contamination: large-scale search identifies more than 2,000,000 contaminated entries in GenBank. Genome Biology 21:115. Conway T, Wazny J, Bromage A, et al. 2012. Xenome - a tool for classifying reads from xenograft samples. Bioinformatics 28(12):i172-i178. Costello M, Fleharty M, Abreu J, et al. 2018. Characterization and remediation of sample index swaps by non-redundant dual indexing. BMC Genomics 19:332. Zhang F, Flickinger M, Taliun SAG, et al. 2020. Ancestry-agnostic estimation of DNA sample contamination from sequence reads. Genome Research 30(2):185-194. Lee S, Lee S, Ouellette S, Park WY, Lee EA, Park PJ. 2017. NGSCheckMate: software for validating sample identity in next-generation sequencing studies within and across data types. Nucleic Acids Research 45(11):e103. Pedersen BS, Bhetariya PJ, Brown J, et al. 2020. Somalier: rapid relatedness estimation for cancer and germline studies using efficient genome sketches. Genome Medicine 12:62.

Related Skills

read-qc/quality-reports - Bimodal GC and overrepresented sequences flag contamination read-qc/adapter-trimming - Remove adapter contamination read-qc/rnaseq-qc - rRNA fraction as a prep-efficiency metric metagenomics/kraken-classification - Deeper taxonomic classification and profiling variant-calling/joint-calling - Where SNP-fingerprint sample swaps do the most damage