mStrain: strain-level identification of Yersinia pestis using metagenomic data

Abstract

Motivation

High-resolution target pathogen detection using metagenomic sequencing data represents a major challenge due to the low concentration of target pathogens in samples. We introduced mStrain, a novel Yesinia pestis strain/lineage-level identification tool that utilizes metagenomic data. mStrain successfully identified Y. pestis at the strain/lineage level by extracting sufficient information regarding single-nucleotide polymorphisms (SNPs), which can therefore be an effective tool for identification and source tracking of Y. pestis based on metagenomic data during a plague outbreak.

Definition

 

Strain-level identification

Assigning the reads in the metagenomic sequencing data to an exactly known or most closely representative Y. pestis strain.

Lineage-level identification

Assigning the reads in the metagenomic sequencing data to a specific lineage on the phylogenetic tree.

canoSNPs

The unique and typical SNPs present in all representative strains.

Ancestor/derived state

An SNP is defined as the ancestor state when consistent with the allele of Yersinia pseudotuberculosis strain IP32953; otherwise, the SNP is defined as the derived state.

Availability and implementation

The code for running mStrain, the test dataset, and instructions for running the code can be found at the following GitHub repository: https://github.com/xwqian1123/mStrain.

1 Introduction

Accurate strain-level target pathogen identification is critical for public health surveillance, especially for controlling disease outbreaks (Buytaers et al. 2021). Metagenomic sequencing is an efficient and advanced technology for the study of microbiomes and target pathogen detection (Quince et al. 2017). The accurate identification of target pathogens in metagenomic data is of major importance. However, the low abundance of target pathogens in a sample complicates their identification. Most state-of-the-art bioinformatics tools, such as MetaPhlAn2 (Truong et al. 2015) and GTDB-Tk (Chaumeil et al. 2019), can only identify pathogens in metagenomic data at the genus or species level (Anyansi et al. 2020). To this end, these tools align sequencing reads against reference genomes of the designated species (Li et al. 2015, Truong et al. 2017, Asnicar et al. 2020). Then the sequence similarity or phylogenetic relationships are analyzed to identify the specific target pathogen (Luo et al. 2015, Wang et al. 2023).

Nevertheless, the sequence alignment methods mentioned above depend on broad databases encompassing numerous species. Such a universal sequence alignment strategy for pinpointing target pathogens at the strain level comes with inherent limitations. First, we need to manually set a sequence similarity cutoff threshold largely based on experience, which is difficult for novice bioinformaticians. Second, we can hardly acquire sequence assemblies due to the low abundance of target pathogens in the test sample. On the other hand, it is common for the presence of highly conserved genomic regions that are shared among different species. These “interspecies repeats” combined with the low coverage of most species, may trigger intergenomic assembly errors (Nurk et al. 2017). Third, the task of aligning raw reads to the comprehensive genome dataset, encompassing reference genomes of multiple species, proves to be a time-intensive endeavor.

Yersinia pestis is the pathogen responsible for plague outbreaks and has caused three pandemics in history (Barbieri et al. 2021). If not treated promptly, Y. pestis infection can cause life-threatening complications. Accurate identification of Y. pestis in metagenomic data is critical for monitoring plague threat. By accurately identifying the strain type of Y. pestis, we can gain valuable insights into the origin of the Y. pestis infection. This knowledge serves as a crucial foundation for clinical treatment, enabling healthcare professionals to tailor their approaches to effectively combat the specific strain involved.

For precise and rapid identification of strain-level target pathogens within metagenomic data, we introduce mStrain (strain-level identification in metagenomic data). In contrast to depending on a comprehensive sequence database that spans multiple species, mStrain focuses solely on the single-nucleotide polymorphism (SNP) loci found within the genomes of the specific target pathogens. These SNPs effectively capture the distinctions among different Y. pestis strains. In this context, we utilize the mStrain methodology to successfully attain strain-level identification of Y. pestis through metagenomic data.

2 Methods

2.1 Framework of mStrain

The framework of mStrain contains two main steps: SNP identification and precise source tracking of target genomes (Fig. 1A). The details of each step to identify Y. pestis in metagenomic data are as follows.

Figure 1.

mStrain pipeline and the details of the identified Y. pestis in the 1400 in silico samples and six mouse blood samples. (A) The mStrain pipeline. (B) The number of in silico samples where Y. pestis was identified to the levels of strain (s), lineage (l), or not-found (non). The horizontal axis is the coverage of the selected genome to the total length of the Y. pestis chromosome genome. The coverage and sequencing depth are both calculated based on the reads belonging to the target pathogen Y. pestis. (C) The number of identified target pathogen SNP loci in the 1400 in silico samples.

For SNP identification of Y. pestis genome sequence in samples, mStrain uses two methods to extract reads belonging to Y. pestis (Fig. 1A II, Method 1 and Method 2). To exclude the reads from close relatives of Y. pestis, we used Kraken2 (v2.0.9) (Wood et al. 2019) to choose the reads belonging to Y. pestis before the alignment (Method 1). In Method 2, mStrain uses the BWA MEM (v0.7.17) aligner (Li 2013) to align the metagenomic raw reads to the Y. pestis reference genome CO92 and extracts aligned reads using SAMtools (v1.9) (Li et al. 2009). It is noted that mStrain aligns the reads to a single reference genome instead of a comprehensive dataset, which helps in reducing the computing time. Both Method 1 and Method 2 are capable of extracting reads that belong to Y. pestis, although they employ slightly different approaches to achieve this. We then use Picard (https://broadinstitute.github.io/picard/) to remove duplicate reads and Trimmomatic (v0.38) (Bolger et al. 2014) to remove reads shorter than 100 bp. Reads were transferred to a .bam file using the BWA-MEM (v0.7.17) aligner (Li 2013). We used BCFtools (v1.9) (Danecek et al. 2021) to filter SNP loci with base quality <30. We then compared filtered SNPs with the prebuilt Y. pestis canoSNP (Cui et al. 2013). CanoSNPs were generated using 134 strains, including 133 Y. pestis representative strains and the ancestor Yersinia pseudotuberculosis strain IP32953. We then removed SNPs with sequencing depths <3× and allele frequencies < 90%. Finally, we identified SNP loci of the Y. pestis strain in metagenomic data. These loci formed an information table with 13 attributes, including position (Pos), chromosome (Chr), reference (REF), allele (A), derived state (Der_St), ancestor state (Anc_St), node (Node), all depth (All_Dpt), filter depth (Fltr_Dpt), depth of the allele identical to the reference base (A_Ref_Dpt), depth of the derived allele (Der_A_St), heterozygote ratio (Het_Ri), variation type (Var_Type), and max-allele frequency number (Max_A_Freq_Num).

In the second step of precise source tracking, we first built a phylogenetic tree based on the canoSNPs using IQ-TREE (v1.6.5) (Nguyen et al. 2015). mStrain then traced back Y. pestis in the metagenomic data based on collected SNPs (Supplementary Table S1). Specifically, mStrain traverses SNPs in the phylogenetic tree structure and determines whether to stop traversing based on SNP state. We first input the root node (root_node) of the phylogenetic tree into the function get_loc and use the function child to obtain child nodes under the root node. We then traverse the child nodes and calculate the number of SNPs in the ancestor states. If the number is not equal to 0, we stop traversing the child node and mark the node on the phylogenetic tree with “X” to indicate that the genotype of Y. pestis may not belong to the branch where the node is located. Otherwise, we continue to traverse the sub-child nodes under the child node. If we can traverse the SNPs to the terminal node, then Y. pestis in the metagenomic data can be identified as the strain shown at the terminal node. Thus, we identify Y. pestis at the strain level. However, if we traverse the SNPs to a middle node, Y. pestis is considered to belong to the lineage where the middle node is located. We then used R software and related packages (Yu et al. 2017, 2018) to draw circular diagrams on each branch of the phylogenetic trees. The circular diagram displays the percentage of the three states of SNPs, including ancestor sate, derived state and not-found, with the number of SNPs of each type in the branch as a proportion of the total number of SNPs in the branch. By tracing back the SNPs using the phylogenetic tree (Supplementary Table S1), mStrain finally identified the target strain in the metagenomic data at the strain level (type s), lineage level (type l), and not-found (type non).

2.2 Datasets used in this study

To verify our mStrain method, we used three different datasets, including 1400 in silico Y. pestis metagenomic data, six mouse blood samples containing low concentrations of Y. pestis, and three clinical samples. Specifically, the 1400 in silico Y. pestis metagenomic data were generated by adding the simulated Y. pestis sequencing reads using ART (Huang et al. 2012) with different sequencing depths and coverages to the clinical sample sequencing reads. The six mouse blood samples were prepared by adding a concentration of 5 × 10²–5 × 10⁴ CFU/ml of Y. pestis to blood samples. Three clinical samples were obtained from the sputum and urine of two patients infected by Y. pestis. More details on the three datasets can be found in the Supplementary Material.

To fully assess mStrain for Y. pestis, we also included two negative control samples containing no Y. pestis strains. Specifically, we first employed a Vibrio fluvialis sample (HY41) as one negative control sample. To further generate the other in silico negative control sample involving off-target close relative Y. pseudotuberculosis, we enhanced the V. fluvialis sample’s sequencing reads by incorporating simulated sequencing reads from Y. pseudotuberculosis strain IP32953. This augmentation involved generating Y. pseudotuberculosis sequencing reads with a depth of 10× and ensuring 100% coverage.

3 Results

We initially validated mStrain using the two negative control samples without and with off-target close relative Y. pseudotuberculosis. In one negative control sample without Y. pseudotuberculosis, mStrain correctly detected the presence of the added V. fluvialis while not identifying any Y. pestis (Supplementary Fig. S3A). In the other negative control sample containing off-target close relative Y. pseudotuberculosis, mStrain accurately identified the presence of Y. pseudotuberculosis instead of Y. pestis (Supplementary Fig. S3B). These findings demonstrate that mStrain can successfully identify the intended target pathogens within the samples.

We used three different datasets to verify mStrain of Y. pestis detection, including in silico Y. pestis metagenomic data, blood samples containing low concentrations of Y. pestis, and clinical samples. mStrain identified Y. pestis at the strain level in each dataset. When processing a compressed FASTQ data sample size of 1.4 GB for the metagenomic sample, mStrain demonstrates a runtime of 4 min and consumes <60.0 GB of memory on a 32-core server. For a larger sample size of 235.0 GB of compressed FASTQ data, mStrain requires 4 h to complete and utilizes <50.0 GB of memory on the same 32-core server.

In the realm of lineage-level identification, mStrain accurately recognized Y. pestis in 1376 out of 1400 simulated metagenomic samples, achieving strain-level discrimination in 1158 of these instances (Fig. 1B, Supplementary Table S2). The number of Y. pestis SNPs identified was positively correlated with sequencing depth and coverage (Fig. 1C). In 90% of metagenomic samples, the lower threshold for identifying Y. pestis at the strain level fell within a range of (10× sequencing depth, 20% coverage) and (5× sequencing depth, 30% coverage) (Fig. 1B). The lower coverage may be compensated for with greater depth, and vice versa. In other words, if the sequencing depth is lower (e.g. 5× depth), then coverage should be higher (e.g. 30% coverage), and vice versa.

In mouse blood samples, mStrain accurately identified Y. pestis at the strain level in one and at the lineage level in two samples. Specifically, mStrain accurately identified the added Y. pestis strain EV76 to the strain level of node 1.ORI3a on the phylogenetic tree, using 191 SNPs, at a Y. pestis concentration of 5 × 10⁴ CFU/ml, 1× depth, and 15% coverage (Supplementary Fig. S1 M1). mStrain identified the added Y. pestis strain EV 76 to the lineage level 1.ORI using only three SNPs, at a Y. pestis concentration of 5 × 10³ CFU/ml, 1× depth, and 4% coverage (Supplementary Fig. S1 M2). Meanwhile, the lineage for the added Y. pestis strain 201 was identified as 0.PE4C, at a concentration of 5 × 10⁴ CFU/ml, 1× depth, and 12% coverage (Supplementary Fig. S1 M4).

In the clinical samples tested, mStrain successfully identified Y. pestis in one patient’s sputum to strain level 2.MED3m (Supplementary Fig. S2 and Table S3), with 1× depth and 38% coverage. Due to the early use of antibiotics, Y. pestis in these clinical samples could not be cultured using traditional methods. Therefore, we were not able to compare our results with whole-genome sequencing after culture. However, our result is consistent with previous Multiple Locus Variable-number tandem repeat Analysis (MLVA) results (Li et al. 2021), which identified the Y. pestis in this plague sample as 2.MED3m.

We evaluated the precision and recall of mStrain by analyzing the three different sample datasets (Supplementary Table S5). For in silico samples, mStrain achieves the precision of 100%/100% and recall of 98.3%/82.7% at the species/strain level. For wet-lab samples, mStrain achieves the precision of 100%/50% and recall of 50%/16.7% at the species/strain level. For clinical samples, mStrain has the precision of 100%/100% and recall of 66.7%/33.3% at the species/strain level.

The average sequencing depth of Y. pestis in wet-lab and clinical samples is typically 1×, below the lower 5× threshold for in silico samples. This low sequencing depth is likely the main reason why identifying Y. pestis in wet-lab and clinical samples is difficult. Despite the challenges posed by low concentration and sequencing depth, mStrain was still able to accurately identify Y. pestis in clinical samples, demonstrating its potential as an effective tool for source tracking of Y. pestis pathogens in metagenomic data at the strain level.

4 Conclusion

Strain-level target pathogen identification is essential for the prevention and control of infectious disease outbreaks. Owing to the low concentration of target pathogens in samples subjected to metagenomic analysis, strain-level identification is a great challenge. Herein, we propose mStrain as an efficient method for strain-level pathogen identification in metagenomic samples, which is based on the alignment between Y. pestis SNPs and canoSNPs. Our results showed that mStrain can accurately identify Y. pestis strains in metagenomic data and clinical samples. Taken together, mStrain can be used as a framework for strain-level target pathogen identification in metagenomic data.

Author contributions

Y.C., R.Y., and X.Z. designed the research; X.Z. and X.Q. wrote the manuscript; X.Q. and Y.W. developed and debugged the code; X.Q., Y.W., and X.Z. analyzed the data. Y.G. and X.P. completed the sequencing of samples from the wet lab experiment. All authors participated in editing the manuscript, and approved the final version.

Supplementary data

Supplementary data are available at Bioinformatics Advances online.

Conflict of interest

None declared.

Funding

This work was supported by the National Key Research and Development Program of China [2022YFC2305304] and the National Natural Science Foundation of China [31970002].