Staphylococcus aureus Whole Genome Sequence-Based Susceptibility and Resistance Prediction Using a Clinically-Amenable Workflow

Abstract

We used graphical user interface (GUI)-based automated analytical tools from Next Gen Diagnostics (Mountain View, California) and 1928 Diagnostics (Gothenburg, Sweden) to analyze whole genome sequence (WGS) data from 102 unique blood culture isolates of Staphylococcus aureus to predict antimicrobial susceptibly, with results compared to those of phenotypic susceptibility testing. Of 916 isolate/antibiotic combinations analyzed using the Next Gen Diagnostics tool, there were 9 discrepancies between WGS-predictions and phenotypic susceptibility/resistance, including 8 for clindamycin and 1 for minocycline. Of 612 isolate/antibiotic combinations analyzed using the 1928 Diagnostics tool, there were 13 discrepancies between WGS-predictions and phenotypic susceptibility/resistance, including 9 for clindamycin, 3 for trimethoprim-sulfamethoxazole and 1 for rifampin. Trimethoprim-sulfamethoxazole was not assessed by Next Gen Diagnostics, and minocycline was not assessed by 1928 Diagnostics. There was complete concordance between phenotypic susceptibility/resistance and genotypic prediction of susceptibility/resistance using both analytical platforms for oxacillin, vancomycin and mupirocin, as well as by the Next Gen Diagnostics analytical tool for levofloxacin (the 1928 Diagnostics tool did not assess levofloxacin). These results suggest that from a performance standpoint, with some caveats, automatic bioinformatics tools may be acceptable to predict susceptibility and resistance to a panel of antibiotics for S. aureus.

Introduction

Our team previously described a whole genome sequencing (WGS) workflow with core genome MLST (cgMLST) analysis of clonality of S. aureus using a graphical user interface (GUI)-based automated analytic system, SeqSphere+ (Ridom, Münster, Germany) (1–3); this method was implemented into our routine clinical practice in 2017. WGS data generated using the same workflow can theoretically be used to predict antibiotic susceptibility and resistance, the so-called “resistance genotype”, “resistome” or “resistotype”; automated analytics for resistance and susceptibility prediction are becoming available and have mainly focused on S. aureus as an initial application.

Several investigators have used S. aureus WGS data to predict antimicrobial susceptibility and resistance; that experience is briefly reviewed here. In 2012, Köser et al. published a small study detailing analysis of WGS data generated on a MiSeq™ (Illumina, San Diego, CA) from 7 neonatal intensive care unit outbreak isolates of methicillin-resistant S. aureus (MRSA) and seven additional MRSA isolates for resistance genes and resistance-associated single nucleotide polymorphisms (SNPs) (4). The investigators’ “resistome” included mecA, ermA, ermC, aacA-aphD, aadD, ant1, tetK, dfrG, fusc, ileS-2, and cfr, alongside manually analyzed chromosomal SNPs compared to antimicrobial resistance-associated mutations in S. aureus proteins detailed in the literature (e.g., Leu466Ser and His481Asn rpoB mutations associated with rifampin resistance, and Ser80Phe grlA and Ser84Leu gyrA mutations associated with ciprofloxacin resistance). While concordance was noted between WGS data analysis and phenotypic susceptibility [determined by the Vitek 2 system version 4.02 (bioMérieux)] to clindamycin, cefoxitin, erythromycin, fusidic acid, gentamicin, kanamycin, mupirocin, tetracycline, trimethoprim, tobramycin, rifampin, ciprofloxacin and linezolid [interpreted according to European Committee on Antimicrobial Susceptibility Testing (EUCAST) 2008 guidelines], their study was limited to 14 MRSA isolates, 7 of which were clonal. Further, they used manual - not automated -, sequence data analysis.

In 2014, Gordon et al. described an analysis of WGS data from 491 S. aureus isolates sequenced on a HiSeq (Illumina) (5). De novo assembled genomes were interrogated using BLASTn against a panel of known resistance-associated chromosomal mutations and genes, with results compared with phenotypic susceptibility testing by automated broth microdilution (BD Phoenix), alongside manual disc diffusion or automated broth microdilution (Vitek), and with selected isolates tested by gradient diffusion for penicillin, methicillin - tested using cefoxitin or oxacillin -, erythromycin, clindamycin, tetracycline, ciprofloxacin, vancomycin, trimethoprim, gentamicin, fusidic acid, rifampin, and mupirocin, and results interpreted using EUCAST breakpoints. There were 60 categorical errors in 48 isolates among 5,193 antimicrobial-isolate pairs tested. There were 25 very major errors, including 6 for ciprofloxacin, 4 for erythromycin, 2 for clindamycin, 4 for fusidic acid, 3 for penicillin, 2 for methicillin, 2 for gentamicin, and 2 for trimethoprim, for an overall very major error rate of 0.5%. There were 35 major errors, 25 of which were for penicillin, for an overall major error rate of 0.7%. The most problematic antimicrobial agent was penicillin; high very major rates were attributed to a variable location of blaZ, on a plasmid or chromosomally integrated, with the latter having coverage which may have been rejected as being of poor quality by the assembly software (5). Major errors were a result of lack of concordance between penicillinase production and disc or broth dilution testing (5). As with the study by Köser et al., an automated analytical tool was not applied.

In 2018, Mason et al. published a study correlating S. aureus WGS data generated using a HiSeq from 1,379 S. aureus isolates. de novo assemblies were analyzed using three publically-available, free, open-source software tools - Mykrobe, Genefinder and Typewriter - with phenotypic susceptibility determined by disc diffusion, automated broth microdilution, and/or agar dilution for ciprofloxacin, clindamycin, erythromycin, fusidic acid, gentamicin, methicillin (predicted), mupirocin, penicillin, rifampin, tetracycline, trimethoprim, and vancomycin, with results interpreted using EUCAST guidelines (6). Overall, 99.5% (113,830/114,457) of individual resistance-determinant/virulence gene predictions were identical between the three WGS analytical methods, with 627 (0.5%) discordances. Genotypic susceptibility or resistance prediction matched phenotype in 98.3% (14,224/14,464). Mykrobe had more false positives for blaZ than the other two pipelines, with Genefinder having a better ability to detect oxazolidinone resistance. There were differences in processing times (i.e., minutes versus hours) between the pipelines studied. Analytic pipelines used were code-level tools, requiring formal bioinformatics support for operation and therefore unsuitable for routine clinical laboratory operation.

In 2019, Babiker et al. published a study correlating phenotypic susceptibility to erythromycin, clindamycin, gentamicin, linezolid, oxacillin, rifampin, trimethoprim-sulfamethoxazole, tetracycline and vancomycin tested by automated broth microdilution with MicroScan WalkAway™ (Siemens Healthcare Diagnostics, Los Angeles, CA) and interpreted using CLSI breakpoints with WGS data generated on a NextSeq (Illumina) for 108 MRSA isolates (7). Resistance genes were assessed using ResFinder, and Comprehensive Antibiotic Resistance Database (CARD) databases. Of 942 isolate/antibiotic combinations, initial concordance between the two methods was 96.7%, with 31 discrepancies between WGS-predicted susceptibility/resistance and automated phenotypic results. Of the 31 discrepant results, 23 resolved by disk diffusion testing, increasing concordance to 99.5%. Following discrepant resolution, WGS-based susceptibility/resistance prediction had an overall sensitivity of 98.5% (95% CI, 96.2–99.6%) and specificity of 99.9% (95% CI, 99.2–100%). Again, the analytic approach used is unsuitable for routine use by clinical laboratory staff.

The SeqSphere+ cgMLST clonality testing method mentioned above is operated in a Windows environment, automated and provides easy to interpret outputs (1–3). In the three years since clinical implementation at our institution, we have found the analytic tool employed for clonality testing to be suitable for use by bench-level laboratory technologists. Although, as detailed above, S. aureus WGS data has been shown to be amenable to prediction of antibiotic susceptibility and resistance, the analytics described above have not been user friendly for bench-level technologists or automated (4–7). Further, the previously described resistance prediction pipelines generally relied on BLASTn queries using de novo assembled genomes requiring significant computational resources and engendering long turnaround times. Here, we used S. aureus WGS data to predict phenotypic susceptibility and resistance of 102 unique blood culture isolates using two GUI-based automated analytics suitable for use in clinical microbiology laboratories, one from Next Gen Diagnostics (Mountain View, California) (8), and the other from 1928 Diagnostics (Gothenburg, Sweden), applying standards that could be used for a new phenotypic susceptibility testing method and with a diverse set of unrelated, including methicillin-resistant and -susceptible, isolates.

Materials and Methods

Bacterial Isolates

102 unique blood culture isolates of S. aureus collected over the course of 2015 at Mayo Clinic in Rochester, Minnesota were studied (2). Isolates were stored at −80°C in MICROBANK^™ freezer vials (Pro-Lab Diagnostics, Richmond Hill, ON, Canada). Archived isolates were recovered and passaged once on trypticase soy agar with 5% sheep blood (BBL, Becton Dickenson, Sparks, MD) prior to DNA extraction and sequencing. Primary recovery culture and subculture plates were incubated in normal atmosphere at 35°C. Seventy-eight isolates and their sequencing have been previously described (2); 24 isolates (SRR5486328, SRR5486313, SRR5486309, SRR5486297, SRR5486295, SRR5486291, SRR5486285, SRR5486274, SRR5486273, SRR5486272, SRR5486271, SRR5486270, SRR5486268, SRR5486267, SRR5486265, SRR5486263, SRR5486262, SRR5486255, SRR5486254, SRR5486253, SRR5486252, SRR5486250, SRR5486249, and SRR5486248) were re-sequenced (and underwent repeat phenotypic susceptibility testing) as a result of discordant phenotypic and genotypic susceptibility on initial testing.

DNA Extraction, Next Generation Sequencing and Data Pre-Processing

Crude DNA was prepared for next generation sequencing by lysing a 4 McFarland standard in tris-EDTA buffer (ThermoFisher, Waltham, MA) with lysostaphin or achromopeptidase (2). 600 μL of cell lysate was transferred to a Maxwell® 16 Tissue DNA Purification Kit cartridge and extracted using the instrument’s on-board protocol (Promega Corporation, Madison, WI) (2). DNA was further purified with the Genomic DNA Clean & Concentrator™ kit (Zymo Research Corporation, Irvine, CA) (2). Sequencing libraries were prepared with NEBNext Ultra™ II (New England BioLabs, Inc., Ipswich, MA) with sequencing performed on a MiSeq using the v2 PE 2×250 chemistry (2). Unaligned BAM files were converted to FASTQ format with samtools (version 1.3.1 10.1093/bioinformatics/btp352). Adapters and indexes were clipped with Trimmomatic (version 0.35 10.1093/bioinformatics/btu170) and then compressed with gzip.

Compressed FASTQ files were uploaded to the GUI-based automated analytics using a ‘drag and drop’ browser interface. Both analytical tools use rapid k-mer based mapping to highly curated resistance databases. Data interrogation and interpretation started automatically.

1928 Diagnostics Software

Samples were uploaded to the 1928 Diagnostics platform where they were subjected to multiple data pre-processing steps to enhance quality of the bioinformatic analysis. This included quality trimming, rejection of samples with an estimated average sequencing depth below 30x, and species identification to rule out sample contamination or mislabeling. Read data was analyzed with an assembly-free k-mer-based method for identification of antibiotic resistance markers (genes and mutations). The proprietary database with resistance markers was constructed from a review of the scientific literature and validated against ~1,000 whole genome sequenced S. aureus isolates with known antibiotic resistance profiles derived from phenotypic susceptibility testing. Database inclusion criteria were that the resistance marker was documented in the scientific literature to confer resistance in the S. aureus, that it was not spuriously associated with resistance due to the underlying population structure and that the minimum inhibitory concentration (MIC) was at or above the clinical breakpoint for resistance specified by the EUCAST (http://www.eucast.org/).

Next Gen Diagnostics Software

The system uploaded FASTQ files for automated processing on a 64-CPU Amazon Web Service cloud instance. Raw reads were trimmed with Trimmomatic v.0.36–4 to remove low quality bases, defined as those with Phred scores <10, from the ends of each read. Further, reads with average scores <20 were assigned “N”. The resulting pool of qualifying reads was assessed by Kraken version 1: https://ccb.jhu.edu/software/kraken with the miniKraken database: https://ccb.jdu.edu/software/kraken/dl/minikraken_20171019_8GB.tgzf for contamination, and then mapped to a S. aureus core genome reference comprising approximately 84% of the ~2.8 Mb S. aureus genomes.

For antimicrobial susceptibility determination, ARIBA (9) was used to compare mapped reads against the system’s resistome database, comprised of resistance genes and gene variants derived from development and validation of a susceptibility prediction model developed using system’s >6,000 sample database of genome/antibiogram pairs. Model development included automated assessment of contribution to sensitivity and specificity of susceptibility prediction for each isolate-antimicrobial pair, assessed for each candidate resistance-conferring gene and mutation, a total of approximately 180,000 test results. The model was iteratively refined by computing the net efficacy of each putative resistance-conferring gene or mutation, examined by removing each in turn and scoring the resulting increase/decrease in sensitivity and specificity, for each antimicrobial agent. Resistance gene inclusion was filtered by thresholds for minimum overlap, identity, and depth required to be considered effective. Thus, model-defined cutoffs were used to filter resistance elements discovered in each mapped genome, imposing requirements, such as minimum coverage to consider a gene present and effective. Grounds for inclusion in the model blended the priority to maximize sensitivity first (to minimize very major errors), followed by optimizing overall accuracy, finally imposing parsimony as a third constraint.

The system furnished predicted susceptibility and resistance and reported each gene or mutation utilized for each resistance prediction, including the model gene or variant discovered, the depth of read coverage, the proportion of the gene covered with qualifying reads, and the degree of match with the resistance element in the database. Resistance and susceptibility were predicted for 31 antimicrobial agents, of which 9 were evaluated in the present work. Susceptibility prediction took <30 seconds per sample.

Phenotypic Antibacterial Susceptibility Testing

Phenotypic susceptibility to oxacillin, clindamycin, rifampin, trimethoprim-sulfamethoxazole, mupirocin (high-level mupirocin screen performed by testing single 256 μg/mL dilution), minocycline, vancomycin, ceftaroline and levofloxacin was determined using agar dilution, and, for daptomycin, by Etest (bioMérieux, Marcy-l’Étoile, France), with results interpreted according to Clinical and Laboratory Standards Institute (CLSI) guidelines (10). Inducible clindamycin resistance was tested by agar dilution using a plate containing erythromycin 4 μg/mL and clindamycin 0.5 μg/mL. The D-test was used for detection of inducible clindamycin resistance to test isolates with discrepant results between phenotypic and genotypic susceptibility testing (10).

Major and Very Major Errors

A very major error was defined as a susceptible genotype with a resistant phenotype, and a major error as a resistant genotype with a susceptible phenotype. Isolates that tested intermediate were excluded from analysis for the affected isolate/antibiotic.

Results

Results are shown in the Table with the full data set shown in the Supplemental Table. Both automatic analytical systems predicted susceptibility or resistance for oxacillin, clindamycin, rifampin, mupirocin and vancomycin. Predictions for trimethoprim-sulfamethoxazole were made solely by 1928 Diagnostics, with predictions for levofloxacin, ceftaroline, daptomycin and minocycline being made only by Next Gen Diagnostics. Notably, 100 (and not 102) isolates were interrogated for minocycline because two isolates tested intermediate.

Table.

Resistance prediction by phenotype versus genotype for the two WGS analytic platforms studied.

Antimicrobial agent	Total no. isolates	No. isolates resistant by phenotype				No. isolates susceptible by phenotype				Very major error rate (%)		Major error rate (%)
		Susceptible by genotype		Resistant by genotype		Susceptible by genotype		Resistant by genotype		Next Gen Diagnostics	1928 Diagnostics	Next Gen Diagnostics	1928 Diagnostics
		Next Gen Diagnostics	1928 Diagnostics	Next Gen Diagnostics	1928 Diagnostics	Next Gen Diagnostics	1928 Diagnostics	Next Gen Diagnostics	1928 Diagnostics
Oxacillin	102	0	0	31	31	71	71	0	0	0	0	0	0
Clindamycin	102	2	3	32	31	62	62	6	6	5.9	8.8	8.8	8.8
Rifampin	102	0	0	1	1	101	100	0	1	0	0	0	1.0
Trimethoprim-sulfamethoxazole	102	Not assessed	0	Not assessed	1	Not assessed	98	Not assessed	3	NA	NA	NA	3.0
Mupirocin2	102	0	0	1	1	101	101	0	0	0	0	0	0
Minocycline	1001	1	Not assessed	0	Not assessed	99	Not assessed	0	Not assessed	100	NA	0	NA
Vancomycin	102	0	0	0	0	102	102	0	0	NA	NA	0	0
Daptomycin	102	0	Not assessed	0	Not assessed	102	Not assessed	0	Not assessed	NA	NA	0	NA
Ceftaroline	102	0	Not assessed	0	Not assessed	102	Not assessed	0	Not assessed	NA	NA	0	NA
Levofloxacin	102	0	Not assessed	32	Not assessed	70	Not assessed	0	Not assessed	0	NA	0	NA
Overall	916/6123	3	3	97	65	810	534	6	10	3.0	4.6	0.8	1.8

¹Two intermediate results were excluded from analysis.

²Mupirocin high-level resistance was either present or not.

³Next Gen Diagnostics/1928 Diagnostics. NA, not applicable

Overall, there were 71 methicillin-resistant and 31 methicillin-susceptible isolates by phenotypic testing; genotypic predictions for oxacillin (methicillin) resistance correlated exactly with phenotypic susceptibility testing for both analytic platforms. Phenotypic testing demonstrated a single high-level mupirocin-resistant isolate and that all isolates were vancomycin-susceptible; genotypic predictions for both mupirocin and vancomycin susceptibility perfectly correlated with phenotypic susceptibility testing for both analytic platforms.

Next Gen Diagnostics yielded 2 very major errors (2/34, 5.9%) and 6 major errors (6/68, 8.8%) for clindamycin, with 1928 Diagnostics yielding 3 very major errors (3/34, 8.8%) and 6 major errors (6/68, 8.8%) for clindamycin. The same 6 isolates yielded major errors by both analytics; ermA was detected in all 6, although they lacked constitutive or inducible clindamycin resistance as determined by an erythromycin-clindamycin combination agar dilution plate and disk diffusion D-zone testing (performed on initial and repeat testing). The associated promoter sequences were consistent with an inducible clindamycin resistance phenotype.

1928 Diagnostics yielded 3 major errors (3/101, 3.0%) in prediction of trimethoprim-sulfamethoxazole susceptibility, with dfrA and dfrG identified in sequencing data from two of three affected isolates. As mentioned above, Next Gen Diagnostics did not assess susceptibility to trimethoprim-sulfamethoxazole. Next Gen Diagnostics yielded a single very major error for minocycline, missing the only phenotypically minocycline resistant isolate. There was a single high-level mupirocin-resistant isolate with mupirocin susceptibility and resistance predicted correctly for all isolates by both analytical tools. There was a single rifampin-resistant isolate, which was called resistant to rifampin by both analytic platforms. 1928 Diagnostics also predicted one rifampin-susceptible isolate to be resistant to rifampin based on identification of a Ser464Pro RpoB mutation.

There were 32 levofloxacin-resistant and 70 levofloxacin-susceptible isolates; only Next Gen Diagnostics assessed for and predicted levofloxacin susceptibility and resistance, doing so with 100% accuracy. All isolates were susceptible to both daptomycin and ceftaroline; susceptibility to both was predicted for all isolates by Next Gen Diagnostics. 1928 Diagnostics did not provide an interpretation for either daptomycin or ceftaroline.

There were 916 antibiotic/isolate combinations analyzed by Next Gen Diagnostics, with an overall very major error rate of 3.0% (3/100) and an overall major error rate of 0.7% (6/816); there were 612 antibiotic/isolate combinations analyzed by 1928 Diagnostics, with an overall very major error rate of 4.4% (3/68) and an overall major error rate of 1.8% (10/544).

Discussion

Our results using WGS data of a limited set of 102 blood culture isolates of S. aureus show low error rates compared to phenotypic susceptibility testing as assessed by agar dilution. Major error rates fell within what could be acceptable limits set by the U.S. Food and Drug Administration for marketing approval of new phenotypic susceptibility testing methods (<1.5% very major error rate, <3% major error rate) for all antibiotics tested except for two. No errors were found for oxacillin, mupirocin, vancomycin, daptomycin (assessed by Next Gen Diagnostics only), ceftaroline (assessed by Next Gen Diagnostics only), or levofloxacin (assessed by Next Gen Diagnostics only). The highest error rates were found with clindamycin; genetic predictions suggest that isolates with very major errors for clindamycin may be inducibly clindamycin resistant, but not detected by current phenotypic methods, although the clinical significance of this finding remains to be determined. Only 1928 Diagnostics assessed for trimethoprim-sulfamethoxazole resistance, with a major error rate of 3%. There was a single rifampin-resistant isolate (Ser464Pro RpoB mutation) which was detected by both systems, with an additional rifampin-susceptible isolate being called rifampin resistant by 1928 Diagnostics (Ser464Pro RpoB mutation). The single minocycline-resistant isolate was missed by Next Gen Diagnostics (minocycline susceptibility was not predicted by 1928 Diagnostics).

A major barrier to clinical adoption of WGS-based methods is a lack of bioinformatics expertise within most clinical microbiology laboratories; our study provides support for the use of automated GUI-based, bioinformatic tools that take WGS data and predict phenotypic susceptibility and resistance to antimicrobial agents, without the need for in-house real-time bioinformatics expertise. Recent publications by Raven et al. and Blane et al. support reliable and reproducible clinical S. aureus sequencing directly from culture plates with a turnaround time from DNA extraction to availability of data files of 24 hours (11, 12). Taken together with our findings and our experience with clinical performance of cgMLST, the turnaround time for identifying a susceptible or resistant genotype in S. aureus is therefore approaching the time required for phenotypic susceptibility testing. As technologies improve and sequencing turnaround time decreases, it may even be possible to predict susceptibility and resistance using WGS data faster than using traditional phenotypic methodologies. Even today though, WGS-based analyses could be helpful for isolates that do not grow well enough to be subjected to phenotypic susceptibility testing, such as small colony variants of staphylococci.

Possible reasons for the observed discrepancies between phenotypic results and WGS-based predictions include errors and variability in phenotypic susceptibility testing, including delayed or lack of expression of resistance mechanisms, and errors in genetic susceptibility testing as a result of loss of resistance elements in subculture, or the presence of novel resistance mechanisms not included in the databases of the analytics used. A challenge for in silico prediction of resistance and susceptibility going forward will be that methods such as those used here will fail to recognize novel variants not in the systems’ databases. In addition, gene expression can be the result of complex interactions between promoters and repressors, and regulatory molecules; regulation may occur at areas remote from resistance genes themselves and therefore be hard to predict, a challenge addressed by phenotypic assays which measure the overall combined impact of all mechanisms. Our study highlights challenges with S. aureus and clindamycin resistance prediction, as highlighted in the past (13, 14). Individual resistance gene coverage may vary depending on whether the gene is located on the chromosome, or if on a plasmid, present in high or low copy number per plasmid and on one or more plasmids carrying the particular gene per isolate. Low coverage regions may be rejected as poor quality because they fall outside the coverage levels of the rest of the genome.

There are several limitations of our study. We did not address prediction of penicillin susceptibility. CLSI suggests that penicillin-susceptible staphylococci should be assessed for inducible β-lactamase before reporting the isolate as penicillin susceptible (10), and that rare isolates that contain genes for β-lactamase production may appear negative by phenotypic tests. In such scenarios, PCR (or conceivably WGS) of the isolate to assess for the blaZ β-lactamase gene may be helpful. Inclusion of penicillin susceptibility testing and assessment for inducible β-lactamase production in future studies would be useful to assess the ability of WGS in penicillin susceptibility prediction. Another limitation is that there was no resistance to vancomycin, ceftaroline, or daptomycin, and low rates of rifampin, trimethoprim-sulfamethoxazole, minocycline and mupirocin resistance in our isolates, so we are unable to confidently evaluate resistance prediction for these agents. In addition, some but not all isolates were tested in duplicate. WGS prediction methods are currently expensive. Finally, our evaluation was limited to S. aureus and cannot be applied to other species.

There are limited publications on the software programs studied. For 1928 Diagnostics, two studies focused on Staphylococcus argenteus (15, 16), with a third analyzing 35 MRSA isolates (all USA 300) in Sweden, several of which were phylogenetically related to one another; genomic resistance traits and the phenotypic resistance profiles were concordant among the investigated isolates (17). There are two publications using the Next Gen Diagnostics platform for prediction of antimicrobial susceptibility/resistance in MRSA. In one, 17 MRSA isolates forming two phylogenetically related clusters (including 9 unique isolates) were studied (18). Phenotypic drug susceptibility to 10 antibiotics was compared with genetic resistance prediction, with full concordance with the exception of fusidic acid susceptibility for one isolate; neither clindamycin nor minocycline, the two most challenging antibiotics for the Next Gen Diagnostics platform in our study, was evaluated (18). In the other, 778 MRSA isolates (including 7 from blood cultures) were studied (8). Overall concordance between phenotypic susceptibility and Next Gen Diagnostics predictions was 99.69%; again, neither clindamycin nor minocycline were studied (8). Overall therefore, ours is the second largest study to date and the only one to compare two automated analytics for susceptibility and resistance prediction from S. aureus WGS data.

We conclude that GUI-based, automated bioinformatics tools may be useful to clinical microbiology laboratories for genotypic resistance prediction in S. aureus.