Abstract
Wastewater surveillance has become essential for monitoring the spread of SARS-CoV-2. The quantification of SARS-CoV-2 RNA in wastewater correlates with the Covid-19 caseload in a community. However, estimating the proportions of different SARS-CoV-2 strains has remained technically difficult. We present a method for estimating the relative proportions of SARS-CoV-2 strains from wastewater samples. The method uses an initial step to remove unlikely strains, imputation of missing nucleotides using the global SARS-CoV-2 phylogeny, and an Expectation-Maximization (EM) algorithm for obtaining maximum likelihood estimates of the proportions of different strains in a sample. Using simulations with a reference database of >3 million SARS-CoV-2 genomes, we show that the estimated proportions accurately reflect the true proportions given sufficiently high sequencing depth and that the phylogenetic imputation is highly accurate and substantially improves the reference database.
Introduction
Wastewater-based epidemiology (WBE) is a vital tool for tracking COVID-19 trends without relying on individual clinical testing. While detecting the virus's presence is straightforward, determining the proportions of specific variants (e.g., Delta vs. Omicron) in a single environmental sample is challenging. Wastewater RNA is often fragmented and sequenced at uneven depths. Furthermore, SARS-CoV-2 strains are genetically similar, often differing by only a few mutations across 30,000 nucleotides. Identifying a specific combination from millions of known genomes presents a massive computational hurdle.
Methodology
The authors developed a three-stage pipeline to address these challenges:
- Filtering: To reduce computational complexity, the method discards candidate strains from the reference database that contain mutations absent in the wastewater sample. This narrows millions of candidates down to approximately 1,000 relevant genomes.
- Phylogenetic Imputation: Public genome databases frequently contain missing data ("N"). Instead of naive "common allele" filling, the authors use the global SARS-CoV-2 phylogeny and a Jukes-Cantor model to predict missing bases. This "Tree Imputation" is significantly more accurate than standard methods.
- EM Algorithm: An Expectation-Maximization algorithm finds the Maximum Likelihood Estimate of strain proportions by iteratively adjusting the assumed percentage of each strain to best explain the observed sequencing reads.
- Likelihood Ratio Test (LLR): To minimize false positives, a threshold of LLR > 4 is used as strong evidence for a strain's presence.
Results
The study utilized a reference database of over 3.1 million SARS-CoV-2 genomes from GISAID. Validation using simulated data with varying error rates and sequencing depths demonstrated high accuracy.
Superior Imputation and Error Correction
Phylogenetic imputation achieved an error rate of < 0.0005, vastly outperforming the "Common Allele" approach, which showed error rates > 0.02 at highly variable sites. Interestingly, this method can "fix" errors in the reference database by identifying true ancestral alleles through the tree structure.
Accurate Proportion Estimation
In simulations involving 10 different strains, the method recovered true proportions with high precision. At a sequencing depth of 1000X, the cumulative error (mass assigned to incorrect strains) was only 0.63%.
Computational Efficiency
By partitioning the 3-million-leaf tree into 121 smaller sub-trees, the authors made the method runnable on standard high-performance computing clusters. A typical sample can be processed in 15 minutes to 2 hours on a single CPU core.
Discussion and Limitations
While robust, the method's accuracy depends on sequencing depth; the authors recommend at least 1000X for reliable surveillance. Detection limits currently focus on strains with a frequency of 5% or higher. Additionally, if two strains are identical across the regions covered by sequencing reads, they are reported as an "unidentifiable group." Future work may address these detection limits and further refine the ability to distinguish nearly identical lineages.
Summary updated