Neptune: A Bioinformatics Tool for Rapid Discovery of Genomic Variation in Bacterial Populations

Eric Marinier, Rahat Zaheer, Chrystal Berry, Kelly Weedmark, Michael Domaratzki, Philip Mabon, Natalie Knox, Aleisha Reimer, Morag Graham, Linda Chui, Gary Van DomselaarbioRxiv 2015 · DOI

TL;DR

Neptune is a new bioinformatics tool designed to efficiently identify genomic regions that are characteristic of specific bacterial populations, even in the presence of genetic variation. It achieves this by using a k-mer matching strategy with probabilistic modeling, enabling rapid and sensitive discovery of these "genomic signatures" without the need for computationally expensive multiple sequence alignments. This makes it particularly valuable for comparative genomics and pathogen discovery.

What Problem Does This Paper Address?

The explosion of genomic sequencing data, especially for bacteria, has outpaced the capabilities of traditional comparative genomics tools. Researchers need to analyze entire populations of bacterial genomes to understand variations that might be linked to specific traits, such as virulence or antibiotic resistance. Existing methods often struggle with the sheer volume of data, are computationally intensive, or are limited in their ability to handle genetic variations within populations. Specifically, there's a need for a tool that can quickly pinpoint larger genomic regions (not just single nucleotide variants) that are present in one group of bacteria but absent in another, while still being robust to minor genetic differences. This is crucial for applications like identifying pathogenic strains or understanding evolutionary adaptations.

Methods

Neptune tackles this problem by employing a k-mer matching strategy with probabilistic modeling. Here's a breakdown of its core approach:

  • k-mer Generation: The software breaks down both "inclusion" (target) and "exclusion" (background) bacterial genomes into short, fixed-length sequences called k-mers. It converts these k-mers to a canonical form (forward or reverse complement) to ensure consistency.
  • Signature Extraction: Neptune identifies "candidate signatures" by looking for regions that are rich in k-mers found in the inclusion group but scarce in the exclusion group. This process allows for a certain number of "k-mer gaps" (regions where expected k-mers are missing) to accommodate natural genetic variation.
  • Probabilistic Modeling: The core of Neptune's novelty lies in its probabilistic model. It uses statistical confidence to determine if a k-mer is sufficiently represented in the inclusion group and sufficiently absent from the exclusion group. This model accounts for base-level mutation probabilities and the GC-content of the environment to estimate the likelihood of k-mer matches.
  • Filtering and Scoring: Candidate signatures are then filtered to remove those that show significant similarity to the exclusion group. The remaining "filtered" signatures are scored based on their sensitivity (how well they match inclusion targets) and specificity (how poorly they match exclusion targets), using a method analogous to BLAST alignments.
  • Parallel Computing: Neptune is designed to leverage parallel computing, allowing it to efficiently process large datasets and extract signatures from draft genome assemblies without requiring computationally expensive multiple sequence alignments.

Datasets

The paper evaluates Neptune using several datasets:

  • Simulated Dataset: A synthetic dataset was created by artificially inserting known genomic islands from Vibrio cholerae into a Bacillus anthracis genome. This allowed for a controlled evaluation of Neptune's ability to identify known regions. Both inclusion and exclusion groups were expanded to 20 genomes each, with a 1% random nucleotide mutation rate introduced.
  • Listeria monocytogenes Dataset: Neptune was applied to identify signatures distinguishing between two clinically relevant serotypes, 1/2a and 4b. This dataset comprised 112 serotype 1/2a and 39 serotype 4b isolates, which were independently tested as inclusion and exclusion groups.
  • Escherichia coli Dataset: Neptune was used to identify signatures specific to Shiga-toxin producing E. coli (STEC) strains. The dataset included 6 STEC (Stx1-producing) and 11 non-STEC draft assemblies.

Key Insights and Ideas

  • "Genomic Signatures" for Population-Level Variation: The paper introduces the concept of "genomic signatures" as regions that are differentially abundant between bacterial populations. This moves beyond single-nucleotide variants to capture larger-scale genomic gains and losses, which are common in bacteria due to horizontal gene transfer.
  • Balancing Sensitivity and Specificity with Probabilistic Models: Neptune's probabilistic approach is key to its success. It allows the tool to be sensitive to variations within a population (e.g., slight differences in gene sequences) while maintaining specificity to the target group, a challenging balance to strike.
  • Efficiency Through k-mer Matching and Parallelization: By avoiding traditional alignment-heavy methods, Neptune offers a significant speedup. The use of k-mers and parallel processing makes it suitable for "big data" genomics.
  • Adaptability to Draft Genomes: The ability to work with draft genome assemblies, which are common for newly sequenced organisms, is a practical advantage.

Key Results

  • High Sensitivity and Specificity: Neptune demonstrated high in silico sensitivity and specificity in all tested datasets. For example, in the Listeria monocytogenes validation, 99.7% of 1/2a signatures aligned to 1/2a targets, and 99.5% of 4b signatures aligned to 4b targets. Crucially, when tested against the other serotype, there were no significant hits, indicating high specificity.
  • Accurate Identification of Known Regions: In the simulated dataset, Neptune correctly identified all six artificially inserted Vibrio cholerae genomic islands, with only one being reported as two adjacent signatures due to a small gap.
  • Discovery of Relevant Signatures: In L. monocytogenes, Neptune identified signatures related to cell wall components, virulence factors, and metabolic pathways, which are biologically relevant to the serotypes. For STEC, it successfully identified the Stx1 toxin-encoding region, as well as other virulence factors.
  • Scalability: The paper demonstrates Neptune's scalability, showing a linear relationship between running time and the number of genomes processed, suggesting it can handle large population studies.

Limitations and Future Work

  • False Positives: Neptune's signature extraction step is designed to avoid false negatives by being sensitive to variations, which can lead to some false positives. These are addressed through filtering, but it increases computational time.
  • Inability to Detect Isolated SNVs: Neptune is designed to find larger genomic regions and cannot detect isolated single nucleotide variants. Regions with high similarity to the exclusion group might be missed.
  • Assumptions in Probabilistic Model: The model assumes probabilistic independence of bases and SNV events, which is a simplification as these events are not truly independent in nature.
  • Short Signatures: Very short signatures (<100 bases) may not be reliably evaluated due to limitations in filtering alignment seed sizes.
  • Redundancy in Consolidated Output: The consolidated signature file, while useful, may contain overlapping signatures, requiring user curation for complete redundancy identification.

Future work could involve further optimization of the filtering process to reduce false positives and exploring ways to incorporate information about specific physical or chemical properties of signatures for application-specific use cases.

Summary updated

Abstract

The ready availability of vast amounts of genomic sequence data has created the need to rethink comparative genomics algorithms using “big data” approaches. Neptune is an efficient system for rapidly locating differentially abundant genomic content in bacterial populations using an exact k -mer matching strategy, while accommodating k -mer mismatches. Neptune’s loci discovery process identifies sequences that are sufficiently common to a group of target sequences and sufficiently absent from non-targets using probabilistic models. Neptune uses parallel computing to efficiently identify and extract these loci from draft genome assemblies without requiring multiple sequence alignments or other computationally expensive comparative sequence analyses. Tests on simulated and real data sets showed that Neptune rapidly identifies regions that are both sensitive and specific. We demonstrate that this system can identify trait-specific loci from different bacterial lineages. Neptune is broadly applicable for comparative bacterial analyses, yet will particularly benefit pathogenomic applications, owing to efficient and sensitive discovery of differentially abundant genomic loci.
BibTeX
@misc{marinier2015neptune,
  author = {Eric Marinier and Rahat Zaheer and Chrystal Berry and Kelly Weedmark and Michael Domaratzki and Philip Mabon and Natalie Knox and Aleisha Reimer and Morag Graham and Linda Chui and Gary Van Domselaar},
  title = {{Neptune: A Bioinformatics Tool for Rapid Discovery of Genomic Variation in Bacterial Populations}},
  year = {2015},
  doi = {10.1101/032227},
  howpublished = {bioRxiv},
}
Figure