PMC

The runtime required by each algorithm as the size of the input increases, as described in the text

On-target classification rate of each method for each confirmed resistance class in the Pediatric and Soil datasets. Alignment and Meta-MARC HTS had improved performance on the Pediatric data than the Soil data, as evidenced by the higher quantiles of sample-wise classification rates shown here for the Pediatric data. Meta-MARC Assembly performed comparably to Resfams. Although alignment and Meta-MARC HTS classified fewer sequence reads overall, their on-target classification rates are comparable to Resfams and Meta-MARC Assembly on the Pediatric dataset

AMR class abundance for the shotgun metagenomic dataset of Noyes et al.. Alignment (green), Meta-MARC HTS (orange), Meta-MARC Assembly (blue) and Resfams (purple) are compared. Meta-MARC Assembly identified the greatest number of reads in 11 out of 13 resistance classes. For the aminoglycosides and multi-drug resistance genes, Resfams identified a greater number of reads with Meta-MARC containing the second largest number of identifications in those classes. Overall, Meta-MARC Assembly detected 42-fold more reads than Alignment, and 1.5-fold more reads than Resfams

Comparison between the total number of classified reads (bar) and the number of on-target classified reads (crosshair) between the various methods. More specifically, Alignment (green), Meta-MARC HTS (orange), Meta-MARC Assembly (blue), and Resfams (purple) are compared for the Pediatric and Soil datasets. The x-axis is labeled by the confirmed AMR class. Of the 388 samples that comprise the Soil and Pediatric test sets, Alignment identified AMR targets in 161 samples (41.5%), Meta-MARC HTS identified 305 samples (78.7%), Meta-MARC Assembly identified 384 samples (98.9%), and Resfams identified 377 samples (97.2%)

Meta-MARC recovers > 90% of simulated reads despite many random point mutations and moderate (25%) contiguous randomization. Percent of genes classified correctly (recovered) is negatively associated with the proportion of the reference sequence randomized before simulation for all methods. Contiguous randomizations (insertion/deletions) demonstrate a consistent negative trend for all methods. Noncontiguous, random point mutations are tolerated well by Meta-MARC Assembly, moderately well for Meta-MARC HTS, and poorly for Alignment and Resfams. Each data point represents a unique annotation Class. The lines were fit using LOESS regression in R, and the gray area is the standard error of the LOESS fit

Meta-MARC utilizes two workflows to classify and count HTS data: Meta-MARC HTS Reads and Meta-MARC Assembly. Meta-MARC HTS Reads Pipeline: HTS reads are input as FASTQ files to be classified by the HMMER software against the pre-built Meta-MARC Models. Resulting counts are processed to correct for multiple classifications; for example, if a single input read is classified to multiple models, the count for that read is divided evenly between the models to maintain a 1:1 input to output ratio. Meta-MARC Assembly Pipeline: HTS reads are de novo assembled to produce contigs. The HTS reads are then aligned back to these assembled contigs to produce an alignment file. The assembled contigs are annotated by HMMER against the Meta-MARC Models. Using the alignment information, HTS reads that also overlap a Meta-MARC model annotation in the assembled contigs are counted. The resulting counts are processed to correct for multiple classifications as described above. The final output of both pipelines is a corrected count file, listing the number of HTS reads classified to each Meta-MARC Model

Comparison between the average number of variations between a read and consensus sequence (contig or reference). Counts of the non-major allele were determined for each genomic position by method and summed by AMR class category and sample. The median value for the methods utilizing assembly and alignment was ~7 variations per read on average. We note that the assembly methods also utilized alignment to map the DNA sequence reads back to the consensus contigs, which contributed to the reduction in allowed variation. The median value for Meta-MARC HTS was threefold higher than competing methods. Meta-MARC HTS was significantly different (***P < 0.001) compared to every other method via a Wilcoxon rank-sum test, corrected for multiple testing by the Bonferroni method. The Wilcoxon rank-sum test was performed using 1891 independent Class-level node and sample combinations per group tested