TL;DR
This paper introduces STRAIN, an R package that uses a novel hybrid approach of read mapping and assembly to accurately assign Multi-Locus Sequence Types (MLST) from whole genome sequencing (WGS) data. STRAIN significantly outperforms existing tools in both identifying correct alleles and discovering novel ones, making it a valuable addition to bacterial typing workflows.
What Problem Does This Paper Address?
Accurately identifying the sequence type (ST) of bacterial isolates is crucial for epidemiological studies, infection control, and understanding bacterial evolution. Multi-Locus Sequence Typing (MLST) is a standard method for this, assigning a unique numerical code based on variations in a set of housekeeping genes. While traditionally done with Sanger sequencing, the increasing availability and decreasing cost of Whole Genome Sequencing (WGS) presents an opportunity to leverage this data for MLST. However, analyzing the short reads generated by WGS for MLST presents a challenge. Existing methods either rely on mapping reads to known alleles (like SRST2), using k-mer analysis (stringMLST), selective assembly (GRAbB), or full genome assembly followed by BLAST (BIGSdb). These methods have limitations, including speed, accuracy in identifying novel alleles, or computational demands. This paper aims to bridge this gap by developing a more accurate and efficient method for MLST from WGS short reads.
Methods
STRAIN employs a hybrid strategy that combines read mapping and assembly. The core idea is to focus on specific MLST loci rather than assembling the entire genome.
- Read Mapping: The process begins by mapping all WGS reads against a database of known MLST allele sequences for each locus using Bowtie2. This step is performed with low stringency to capture a broad range of potential matches.
- Read Filtering: The mapped reads are then filtered into four sets based on increasing alignment stringency (all reads, reads matching at least 30%, 60%, and 90% of their length). This aims to improve the quality of reads used for assembly.
- Locus-wise Assembly: Each filtered read set for each locus is then independently assembled using SPAdes. As a backup, Velvet is used if SPAdes fails. This "reduced assembly" approach focuses computational resources on the relevant regions.
- Allele Identification: The assembled contigs are then aligned using BLASTn against the reference allele database. A perfect match identifies the correct allele. If no perfect match is found, the program flags it as a potential "new allele" and stores its sequence for further investigation.
This hybrid approach aims to leverage the speed of mapping while retaining the accuracy of assembly for identifying novel variants.
Datasets
The primary dataset used for evaluation was a benchmark dataset of 540 publicly accessible Illumina read sets, originally published by Inouye et al. (2014). This dataset was chosen because it allowed for direct comparison with previously published tools. The STs and allele assignments determined by PCR or capillary sequencing in the original study were considered the "correct" ones for evaluation. The authors also created modified reference datasets by removing the correct alleles to specifically test the tools' ability to identify novel alleles.
Key Insights and Ideas
- Hybrid Approach Synergy: The combination of mapping and assembly, applied in a locus-specific manner, appears to be more powerful than either method alone. Mapping helps to quickly identify relevant reads, while assembly allows for the reconstruction of potentially novel allele sequences.
- Focusing on Loci: By not requiring a full genome assembly, STRAIN significantly reduces computational time and resource requirements, making it more accessible.
- Novelty Detection: The ability to not only identify correct alleles but also to flag and extract sequences of novel alleles is a critical advantage for tracking bacterial evolution and identifying emerging strains.
- R Package Convenience: Implementing STRAIN as an R package makes it easily integrable into existing bioinformatics pipelines and accessible to a broad user base.
Key Results
STRAIN demonstrated superior performance across several metrics when compared to SRST2, stringMLST, and GRAbB on the benchmark dataset:
- Correct Allele Identification: STRAIN achieved a 97.0% capability to identify correct alleles, correctly assigning 3666 out of 3780 alleles. This was significantly higher than other tools, which ranged from 28.5% to 96.9%.
- New Allele Identification: STRAIN showed an impressive 98.7% capability to identify new alleles, successfully identifying them in 3730 out of 3780 cases. This also surpassed other methods, which ranged from 0% to 97.1%.
- Reduced No-Output: STRAIN had a low percentage of no-output (0.6%), indicating robust performance across diverse samples.
- Speed: While not the primary focus, the paper notes that STRAIN's parallelized implementation in R, especially with multi-threading, drastically reduces running time, with a median of 8 minutes and 13 seconds for complex cases when run in parallel.
Limitations and Future Work
- Dependence on External Tools: STRAIN relies on several external programs (Bowtie2, SPAdes, Velvet, BLASTn). While these are common, their installation and compatibility could be a minor hurdle for some users.
- Potential for Novelty Misclassification: While STRAIN excels at identifying novel alleles, the paper acknowledges that distinguishing between a truly novel allele and a sequencing artifact or a very rare known allele can still be challenging.
- Future Work: The authors suggest that STRAIN's locus-wise approach makes it adaptable for typing other genetic loci beyond MLST, such as drug resistance genes, ribosomal genes, or antigens, provided appropriate reference databases are available. They also highlight its potential for integration into broader downstream bioinformatics analyses.
Summary updated