Contamination Screening - Usage Guide

Overview

Contamination comes in five distinct classes, each with its own detection method and fix. A species screen (FastQ Screen, Kraken2) maps reads to genome references and answers "what organisms are here?" -- but it is structurally blind to same-species cross-sample contamination and sample swaps, which need SNP fingerprinting (verifyBamID2, NGSCheckMate, somalier, conpair). Any human or single-species cohort needs both. The default action is to report contamination, not filter, because removing reads biases composition.

Prerequisites

conda install -c bioconda fastq-screen kraken2 bracken bbmap sortmerna somalier
# Pre-built FastQ Screen databases
fastq_screen --get_genomes

Quick Start

Tell your AI agent what you want to do: - "Screen my FASTQ files for cross-species contamination" - "Check if my human sample has mouse (PDX host) contamination" - "Verify my human samples are not swapped or mixed" - "Estimate the contamination fraction in my BAM"

Example Prompts

Cross-species screening

"Run FastQ Screen on my samples against human, mouse, and common contaminants"

"Classify my reads with Kraken2 to find a bacterial contaminant"

Same-species (swaps and mixtures)

"My RNA-seq results look wrong; check whether samples are swapped"

"Estimate within-species contamination in my tumor and normal BAMs"

PDX and remediation

"Separate human graft from mouse host reads in my PDX sample"

"Remove PhiX spike-in from my reads before assembly"

What the Agent Will Do

Identify which contamination class is in question (cross-species, swap, vector, rRNA, cell-line)
Run the matching detector (genome panel for organisms; SNP fingerprints for individuals)
Read the FastQ Screen bar chart by category (unexpected one-hit-one-genome = contamination; multiple-genomes = homology)
Recommend report vs filter, preferring combined-reference alignment for PDX
For human cohorts, add a SNP-fingerprint identity/contamination check

Interpretation Guide

FastQ Screen bar chart (cross-species only)

An unexpected genome with high One_hit_one_genome = real foreign DNA.
Reads spread across multiple genomes' multiple-hits categories = conserved/homologous sequence (rRNA, mito), not contamination.
High Hit_no_genomes = adapter dimer, a missing reference, or a novel organism.

Same-species signals (NOT visible to FastQ Screen)

verifyBamID2 FREEMIX > ~0.02 = within-species contamination.
FREEMIX near 0 but CHIPMIX near 1 = a sample swap, not contamination.
somalier/NGSCheckMate off-diagonal identity = swapped or mixed samples.

Tips

Run an organism screen early, before investing in alignment.
A species screen cannot detect a human-A/human-B swap; use SNP fingerprints for that.
Index hopping on patterned flowcells creates phantom low-VAF variants; use unique dual indexing for ctDNA/single-cell/somatic.
Default to report-only; filter only a NAMED contaminant (PhiX, adapters) with a precise k-mer remover (BBDuk).
For PDX, prefer aligning to a combined human+mouse reference over hard pre-filtering.
Reference-genome contamination can corrupt the screen; trust a curated database and treat surprising single-source hits as artifacts until ruled out.

Resources

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