Phylogeny-aided detection of contamination in nearly 5 million SARS-CoV-2 genomes

Technical Summary: Phylogeny-aided Detection of Contamination in Nearly 5 Million SARS-CoV-2 Genomes

Problem Statement
The massive scale of SARS-CoV-2 genome sequencing, exceeding 17 million sequences in public repositories, has introduced significant challenges in data quality. A primary source of error is sample contamination, where multiple genomes from different hosts are accidentally mixed during sequencing. While high-abundance contaminants are often detectable, low-abundance contamination combined with "amplicon dropout" presents a specific, difficult-to-detect issue. In amplicon-based sequencing (the standard for SARS-CoV-2), mutations at primer binding sites can prevent the amplification of the majority genome in a contaminated sample. Consequently, sequencing reads from the minority contaminant genome may dominate specific regions, leading to a hybrid consensus sequence that falsely appears as recombination or introduces erroneous mutations. These artifacts can disrupt downstream analyses, including lineage assignment, transmission tracking, and evolutionary inference. Existing computational methods often fail to account for the specific interplay between low-level contamination and amplicon dropout.

Methodology: The PhyCD Pipeline
The authors present PhyCD (Phylogeny-aided Contamination Detection), a computational pipeline designed to identify and mask consensus sequence errors resulting from the combination of contamination and amplicon dropout. The pipeline operates through the following steps:

1. Data Preprocessing and Reference Selection:

   • The study utilizes a dataset of 4,952,451 SARS-CoV-2 genomes assembled using Viridian, which already addresses reference biases and masks low-coverage or highly heterozygous sites.
   • A "reference set" of 785,011 samples is selected based on strict criteria: they must have zero positions below a specific coverage threshold and very few heterozygous sites. These samples are assumed to be free of contamination-related consensus errors.
   • A pandemic-scale reference phylogenetic tree is inferred from this reference set using MAPLE (v0.7.5).

2. Dropout Masking:

   • For the remaining "tested" samples, the pipeline identifies regions where sequencing coverage drops significantly relative to the sample's median coverage.
   • A threshold $\rho$ (tested at 5%, 10%, 15%, 20%) is applied. Positions with coverage below $\max(50, \rho \times M_s)$ (where $M_s$ is the median coverage) are masked. This targets regions where the majority genome's reads may have dropped out, leaving only contaminant reads.
   • For each sample, three consensus sequences are generated:
     • Dropout masked: Regions masked based on coverage drops.
     • Dropout unmasked: The standard consensus (used as a baseline).
     • Randomly masked: A control where the same number of positions are masked randomly to distinguish true signal from noise.

3. Phylogenetic Placement and Signal Detection:

   • All three sequence types are phylogenetically placed onto the reference tree using MAPLE.
   • The core logic relies on the assumption that contamination-induced errors create artificial substitutions that increase the phylogenetic distance (branch length) from the true evolutionary lineage.
   • If the dropout masked sequence places significantly closer to the reference tree (shorter branch length) than the dropout unmasked sequence, it suggests the masked regions contained erroneous substitutions caused by contamination.
   • Samples where the dropout masked sequence is at least 2 substitutions closer to the tree than the unmasked sequence are flagged as "putative contaminated."

4. Contaminant Search (Optional):

   • The pipeline attempts to identify "plausible contaminants" among the reference samples by checking for similarity in the masked regions and heterozygous sites. However, the authors note that likelihood-based models and closeness scores did not achieve sufficient discriminatory power to filter out false positives effectively in this specific dataset.

Key Results

• Detection Scale: Applied to nearly 5 million genomes, PhyCD identified 10,942 putative contaminated samples using conservative parameters ($\rho = 5\%$).
• False Positive Estimation: By comparing against the randomly masked control, the authors identified 5,457 samples flagged by the random masking process. This suggests that approximately 50% of the flagged samples may be false positives (i.e., genuine mutations masked by the coverage drop).
• Impact on Data: Across the flagged genomes, PhyCD identified and permitted the masking of 64,753 substitutions that could otherwise cause errors in downstream analyses.
• Parameter Sensitivity: Increasing the coverage threshold $\rho$ increased the number of flagged samples but degraded the signal-to-noise ratio (from 2.0 at $\rho=5\%$ to ~1.6 at $\rho=20\%$). Higher thresholds led to excessive loss of genuine genetic information.
• Case Study: Analysis of a specific sample (SRR24221569) demonstrated that masking coverage-drop regions reduced the phylogenetic placement branch length from 2.1 to 0, effectively placing the sample directly onto an existing reference sample and increasing placement confidence from 0.32 to 0.94. Read-depth visualization confirmed the presence of a coverage drop and heterozygosity consistent with the contamination hypothesis.

Significance and Claims
The paper claims that PhyCD provides a scalable, phylogeny-aided method to detect contamination artifacts that are invisible to standard heterozygosity-based filters due to amplicon dropout.

• Trade-off Management: The authors emphasize the inherent trade-off between sensitivity and precision. While a 50% false positive rate might seem high, the authors argue that in the context of global surveillance, it is advantageous to mask or exclude ~5,000 potentially contaminated samples (0.1% of the total dataset) to prevent the propagation of artifacts in recombination inference and lineage assignment.
• Scalability: The approach successfully handles pandemic-scale data (millions of genomes) by leveraging efficient phylogenetic placement tools like MAPLE.
• Generalizability: While developed for SARS-CoV-2, the authors suggest the method could be extended to other amplicon-sequenced pathogens, provided a sufficiently dense reference phylogeny exists.
• Limitations: The authors modestly acknowledge the lack of a "gold-standard" dataset of experimentally verified contaminated samples to validate true positive rates. They also note that the current pipeline relies on sequential filtering steps rather than a unified probabilistic model, which could theoretically offer higher power but presents computational challenges at this scale.

In conclusion, the study demonstrates that phylogenetic placement can serve as a reliable signal for contamination detection at a pandemic scale, revealing that while contamination affecting consensus sequences is relatively rare (~0.1%), the absolute number of affected genomes is significant enough to warrant automated curation tools like PhyCD.