Hierarchical Hidden Markov models enable accurate and diverse detection of antimicrobial resistance sequences

Steven M. Lakin, Alan Kuhnle, Bahar Alipanahi, Noelle Noyes, Chris Dean, Martin D. Muggli, Rob Raymond, Zaid Abdo, Mattia Prosperi, K. E. Belk, Paul S. Morley, Christina BoucherCommunications 2019 · DOI

Abstract

The characterization of antimicrobial resistance (AMR) genes from high-throughput sequencing data is foundational in public health research. Traditional alignment methods, while specific, often fail to detect sequences divergent from reference databases, resulting in high false-negative rates. We introduce Meta-MARC, which utilizes hierarchical, DNA-based Hidden Markov Models (HMMs) to enable the detection of diverse resistance sequences. Meta-MARC demonstrates higher sensitivity than competing methods, allowing for the discovery of divergent AMR genes and the expansion of existing databases.

Introduction: The "Dark Matter" of Antibiotic Resistance

The global rise of antimicrobial resistance is a public health crisis. To track the spread of resistance, researchers sequence DNA from various environments—the "resistome." Traditionally, identifying AMR genes relies on alignment: matching short sequence reads against a database. However, alignment is brittle. If a gene has evolved even slightly, tools like BWA often fail to map it. This leaves researchers blind to novel or divergent resistance threats. While protein-based models exist, they typically require DNA assembly and translation, which are error-prone and computationally expensive.

Methodology: The Meta-MARC Framework

Meta-MARC addresses these challenges by employing Hierarchical Hidden Markov Models (HMMs). Unlike alignment's character-for-character matching, HMMs model the probabilistic "profile" of a gene family, capturing the underlying biological signal even when the specific sequence varies.

Meta-MARC workflow showing the HTS reads and assembly pipelines.

Key technical innovations include:

  • DNA-Based Modeling: Meta-MARC works directly on raw DNA reads, bypassing the need for sequence assembly—a major bottleneck for low-abundance samples.
  • Hierarchical Structure: It mirrors the MEGARes database, classifying sequences across four levels: Class, Mechanism, Group, and Model. This provides high-confidence broad classifications even for unknown variants.
  • Clustering Strategy: Models were built by clustering nearly 3,000 sequences at 80% nucleotide identity, ensuring representation of diverse gene families.

Results: Mapping the Resistome

The researchers evaluated Meta-MARC using simulated data, functional metagenomics (soil and pediatric), and shotgun metagenomics from beef cattle production.

Comparison of AMR read identification across different methods in pediatric and soil datasets.

Meta-MARC demonstrated superior detection capabilities:

  • Increased Sensitivity: In shotgun metagenomic data, Meta-MARC identified 42 times more resistance sequences than alignment and 1.5 times more than Resfams (a leading protein-HMM tool).
  • High Accuracy: On simulated data, the tool achieved >98% sensitivity and >99% specificity.
  • Discovery Potential: Meta-MARC identified resistance in 11 out of 13 classes where alignment found nothing, effectively mapping the "dark matter" of the cattle resistome.
  • Robustness: The tool can tolerate approximately 22 more mismatches per read than alignment, making it uniquely suited for non-model environments like soil.

Discussion and Limitations

While Meta-MARC dramatically increases the "search radius" for divergent genes, it comes with a computational cost. The raw read pipeline is significantly slower than alignment, and the memory footprint is substantial (requiring ~324 GB RAM).

However, the core finding is that high specificity can be maintained while significantly increasing sensitivity. Future work suggests an integrated pipeline: using fast alignment for "known" genes and Meta-MARC for the remaining "unmapped" reads to balance speed and discovery. As AMR databases evolve, Meta-MARC's HMMs will continue to provide a robust framework for expanding our understanding of global resistance.

Summary updated

Abstract

The characterization of antimicrobial resistance genes from high-throughput sequencing data has become foundational in public health research and regulation. This requires mapping sequence reads to databases of known antimicrobial resistance genes to determine the genes present in the sample. Mapping sequence reads to known genes is traditionally accomplished using alignment. Alignment methods have high specificity but are limited in their ability to detect sequences that are divergent from the reference database, which can result in a substantial false negative rate. We address this shortcoming through the creation of Meta-MARC, which enables detection of diverse resistance sequences using hierarchical, DNA-based Hidden Markov Models. We first describe Meta-MARC and then demonstrate its efficacy on simulated and functional metagenomic datasets. Meta-MARC has higher sensitivity relative to competing methods. This sensitivity allows for detection of sequences that are divergent from known antimicrobial resistance genes. This functionality is imperative to expanding existing antimicrobial gene databases.
BibTeX
@inproceedings{lakin2019hierarchical,
  author = {Steven M. Lakin and Alan Kuhnle and Bahar Alipanahi and Noelle Noyes and Chris Dean and Martin D. Muggli and Rob Raymond and Zaid Abdo and Mattia Prosperi and K. E. Belk and Paul S. Morley and Christina Boucher},
  title = {{Hierarchical Hidden Markov models enable accurate and diverse detection of antimicrobial resistance sequences}},
  booktitle = {Communications},
  year = {2019},
  doi = {10.1038/s42003-019-0545-9},
}

Loading related content...

Figure