Taking the next-gen step: Comprehensive antimicrobial resistance detection from Burkholderia pseudomallei

Abstract

Background

Antimicrobial resistance (AMR) poses a major threat to human health. Whole-genome sequencing holds great potential for AMR identification; however, there remain major gaps in accurately and comprehensively detecting AMR across the spectrum of AMR-conferring determinants and pathogens.

Methods

Using 16 wild-type Burkholderia pseudomallei and 25 with acquired AMR, we first assessed the performance of existing AMR software (ARIBA, CARD, ResFinder, and AMRFinderPlus) for detecting clinically relevant AMR in this pathogen. B. pseudomallei was chosen due to limited treatment options, high fatality rate, and AMR caused exclusively by chromosomal mutation (i.e. single-nucleotide polymorphisms [SNPs], insertions-deletions [indels], copy-number variations [CNVs], inversions, and functional gene loss). Due to poor performance with existing tools, we developed ARDaP (Antimicrobial Resistance Detection and Prediction) to identify the spectrum of AMR-conferring determinants in B. pseudomallei.

Findings

CARD, ResFinder, and AMRFinderPlus failed to identify any clinically-relevant AMR in B. pseudomallei; ARIBA identified AMR encoded by SNPs and indels that were manually added to its database. However, none of these tools identified CNV, inversion, or gene loss determinants, and ARIBA could not differentiate AMR determinants from natural genetic variation. In contrast, ARDaP accurately detected all SNP, indel, CNV, inversion, and gene loss AMR determinants described in B. pseudomallei (n≈50). Additionally, ARDaP accurately predicted three previously undescribed determinants. In mixed strain data, ARDaP identified AMR to as low as ~5% allelic frequency.

Interpretation

Existing AMR software packages are inadequate for chromosomal AMR detection due to an inability to detect resistance conferred by CNVs, inversions, and functional gene loss. ARDaP overcomes these major shortcomings. Further, ARDaP enables AMR prediction from mixed sequence data down to 5% allelic frequency, and can differentiate natural genetic variation from AMR determinants. ARDaP databases can be constructed for any microbial species of interest for comprehensive AMR detection.

Funding

National Health and Medical Research Council (BJC, EPP, DSS); Australian Government (DEM, ES); Advance Queensland (EPP, DSS).

Research in context.

Evidence before this study

If unchecked, antimicrobial resistance (AMR) is predicted to have a devastating impact on global health in the coming decades. Next-generation sequencing (NGS) is an essential tool for combatting AMR, providing a comprehensive and accurate diagnostic tool for AMR detection and unveiling the molecular basis underpinning the evolution of AMR in many dangerous multidrug-resistant pathogens. Whilst currently available AMR software readily detects horizontally-acquired AMR genes and some chromosomally-encoded variants, no existing tool can detect AMR determinants caused by the spectrum of chromosomal mutations, leading to considerable underreporting of AMR in many microbes.

Added value of this study

To overcome current software limitations, we were prompted to develop ARDaP. Using NGS or genome assembly data as input, ARDaP can detect and predict AMR caused by gene acquisition, point mutations, insertions-deletions, gene copy-number variation, inversions, and gene loss or truncation. We tailored ARDaP for AMR determinant detection in the formidable melioidosis pathogen, Burkholderia pseudomallei, which has limited treatment options due to intrinsic multidrug resistance and poor or no AMR detection support with existing AMR software. ARDaP also incorporates a mixture-aware feature that enables the detection of emerging AMR determinants, thereby informing early treatment shifts and improving antibiotic stewardship efforts and patient survival. Although we demonstrate its application in B. pseudomallei, ARDaP databases can be developed to identify AMR in any microbe of interest.

Implications of all the available evidence

Using ARDaP, both known and novel AMR determinants can be accurately identified from NGS data, and non-AMR-conferring variants can be ignored, representing important advances over existing AMR detection software. Inclusion of antimicrobial-susceptible strains, an important yet often-overlooked component of AMR database development and validation, is critical for accommodating natural genetic variation and mitigating high false-positive rates. Functional verification of novel AMR determinants (e.g. phenotypic testing, gene knockouts, heterologous expression, or RNA sequencing), remains a limiting factor in our understanding of AMR. Our study highlights the essential need for well-curated and meticulous pathogen-specific databases for the most accurate, comprehensive, and clinically relevant AMR detection. Ongoing efforts are needed to continue uncovering the myriad ways that microorganisms evolve to evade antimicrobial agents.

Alt-text: Unlabelled box

1. Introduction

Antimicrobial resistance (AMR) poses a major threat to human health worldwide and is an increasing contributor to morbidity and mortality. Antibiotic use and misuse have resulted in an alarming increase in multidrug-resistant infections worldwide, provoking an urgent need to improve global AMR detection and surveillance. Alongside pathogen identification, AMR detection is one of the primary goals of diagnostic microbiology, with far-reaching consequences for both infection control and effective treatment [1].

Whole-genome sequencing (WGS) permits comprehensive AMR detection and prediction from bacterial genomes by identifying all AMR determinants in a single genome or metagenome [2], circumventing the need for multiple and often laborious diagnostic methods. Existing bioinformatic tools such as ARG-ANNOT [3], Antibiotic Resistance Identification By Assembly (ARIBA) [4], Comprehensive Antibiotic Resistance Database (CARD) [5], ResFinder [6], AMRFinder [7], and MEGARes [2] can readily detect AMR genes acquired from horizontal gene transfer events. Many bacterial pathogens also develop AMR via chromosomal mutations, including missense single-nucleotide polymorphism (SNP) mutations in β-lactamase-encoding genes, SNPs or insertion-deletions (indels) in efflux pump regulators [8], [9], [10], gene amplification via copy-number variations (CNVs) [11], inversions [9], and functional gene loss [8]. Recent improvements in AMR identification software mean that chromosomal mutations, particularly SNPs, are now identifiable. For example, ARIBA can identify AMR-conferring SNPs and indels in multiple species [4]. Nevertheless, other types of genetic variants – gene loss or truncation, inversions, and CNVs – remain poorly identified using existing tools, despite their crucial role in conferring AMR [12].

The Tier 1 Select Agent bacterium, Burkholderia pseudomallei, causes the often-fatal tropical disease melioidosis. Melioidosis severity ranges from mild, self-limiting skin abscesses to pneumonia, neurological disease, and septic shock. B. pseudomallei is naturally resistant to many antibiotics, including aminoglycosides, penicillins, macrolides, and polymyxins [13,14]. Fortunately, human-to-human B. pseudomallei transmission is rare; almost all infections are acquired from the environment. As such, isolates collected prior to antibiotic treatment are almost universally susceptible to the following clinically-relevant antibiotics: ceftazidime (CAZ), amoxicillin-clavulanate (AMC), co-trimoxazole (SXT), doxycycline (DOX), meropenem (MEM) and imipenem (IPM) [15]. To prevent melioidosis relapse, treatment involves prolonged (3-6 month) antibiotic therapy, which increases AMR risk and treatment failure [8]. AMR in B. pseudomallei has been reported for all clinically-relevant antibiotics [8], with novel AMR determinants towards these key antibiotics continuing to be uncovered.

Here, we tested 47 characterised B. pseudomallei genomes with known antibiotic phenotype profiles and associated AMR determinants, and three MEM-resistant (MEMr) strains with previously unidentified AMR determinants, against existing tools (ARIBA, CARD and AMRFinderPlus) to determine their AMR detection efficacy. Among the characterised strains, 25 were phenotypically-confirmed as resistant towards at least one clinically relevant antibiotic, 16 were sensitive, and the remainder encoded unusual sensitivity towards aminoglycosides and macrolides, or stepwise AMR variants. Following testing against the current AMR tools, we developed a new tool, Antibiotic Resistance Detection and Prediction (ARDaP), to permit comprehensive AMR detection from microbial genomes. ARDaP was designed to meet four main aims: first, to accurately identify AMR determinants caused by a spectrum of mutational mechanisms (i.e. gene gain, SNPs, indels, CNVs, inversions, and functional gene loss); second, to predict enigmatic AMR determinants in isolates with phenotypically-confirmed AMR, third, to detect minor AMR allelic determinants in mixed (e.g. metagenomic) sequence data; and finally, to provide a user-friendly report that summarises the AMR determinants (if any) and associated AMR phenotypes, stepwise variants, unusual antimicrobial sensitivity determinants, and genetic variants associated with natural variation that do not confer AMR. Although we illustrate its utility in B. pseudomallei, ARDaP is amenable to AMR identification across all microbial species.

2. Methods

Ethics. Ethics approval was obtained from the Human Research Ethics Committee of the Northern Territory Department of Health and Menzies School of Health Research (HREC 02/38, “Clinical and Epidemiological Features of Melioidosis”). Written informed consent was provided by study participants.

Isolates. Forty-seven B. pseudomallei strains were included in this study, including 25 with elevated MICs towards one or more clinically-relevant antibiotics (Table 1) and genotypically-confirmed AMR determinants. These isolates were selected as they represent the spectrum of known AMR determinants in B. pseudomallei (Table S1). Strains encoding unusual aminoglycoside- and macrolide-sensitivity, and stepwise mutations that lower the barrier to AMR development, were also examined (Table 1). A further 16 strains sensitive to all clinically-relevant antibiotics were included to test software efficacy (Table 2). Finally, three previously uncharacterised clinical strains exhibiting MEMr (MSHR1058 MIC=12 µg/mL; MSHR1174 MIC=6 µg/mL; MSHR8777 MIC=4 µg/mL; Table 3) were included to test the predictive capacity of ARDaP.

Table 1.

Antimicrobial resistance (AMR) determinants in 25 Burkholderia pseudomallei strains with verified AMR phenotypes, plus strains conferring unusual antimicrobial susceptibility and stepwise AMR variants^§.

Patient ID	Isolate	ST	Genome accession	Antibiotic MIC (µg/mL)¥	Stepwise variant†	AMR determinant	Reference/s
Thai patient	316c	17	SRR2975745	CAZr (64)	—	PenAP173S	[22,23]
Thai patient	354e	78	AHJD00000000.1	CAZi (6)	—	penA -78G>A	[9]
				SXTi (3)	BpeT structural variant‡	Ptr1R21fs
Australian patient	Bp1651	880	SRR2102060	CAZr (≥128)	penA -78G>A	PenAD245G	[38]
				Unusual sensitivity to aminoglycosides (GEN, KAN) and macrolides (AZM)	—	AmrBA254fs
				DOXr (16)	—	BPSL3085A88fs
				AMCr (64/32)	penA -78G>A	PenAS78F
Pre-DPMS 89	MSHR0052	722	SRR5818275	MEMr (8)	—	AmrRE190*	[8]
				DOXr (48)	AmrRE190*	BPSL3085V211M
Australian patient	MSHR0292	236	SRR4254580	DOXr (16)	AmrRS174P	BPSL3085V40A	[29]
P215	MSHR0663	36	SRR2887062	SXTr (≥32)	BpeTH278Y	Ptr1R21fs	[8]; This study
	MSHR0937	36	SRR2886988	AMCi (12/6), MEMr (6)	—	BpeRD176A
				SXTi (3)	BpeRD176A	MetFQ142*
P179	MSHR0678	114	SRR6075118	MEMr (3)	—	AmrRE21D	[8]
	MSHR0800	114	SRR6075115	MEMr (6), DOXi (8)	—	BpeRL85fs
P337	MSHR1226	333	SRR9598635	CAZr (≥256)	—	PenAC75Y	[22]
	MSHR1300	333	SRR6075114	CAZr (≥256)	penA -78G>A	PenAC75Y	[22]
				MEMr (4)	—	AmrRK13fs	[8]
				SXTi (3)	—	AmrRK13fs	This study
P595	MSHR3683	144	SRR11678542	DOXi (12)	—	BPSL3085A88fs	This study
P608	MSHR4083	36	SRR2887030	SXTr (24)	AmrRΔA153-D156	DutG91A	[8]
				MEMr (6)	—	AmrRΔA153-D156
CF6	MSHR5654	1040	SRR3404570	CAZr (≥256)	PenA 30x CNV	PenAC75Y	[11]
				CIPr (≥32)	—	GyrAY77S
				SXTr (≥32)	BpeTT314fs	Ptr1R21fs
CF9	MSHR5665	252	SRR3404582	SXTr (≥32)	Ptr1W116R; MetFQ28P	DutV77A	[11,25]
				DOXi (6)	—	BPSL3085A88fs
	MSHR5666	252	SRR3404597	DOXs (1)#	—	BPSL3085A88fs
				SXTr (≥32)	Ptr1W116R; MetFQ28P	DutV77A
	MSHR5667	252	SRR3404598	SXTr (≥32)	MetFN162T; AmrRL132P	DutN99S
				MEMr (4)	—	AmrRL132P
				DOXr (48)	AmrRL132P	BPSL3085A88fs
	MSHR5669	252	SRR3404599	DOXi (6)	—	BPSL3085A88fs
				SXTr (≥32)	Ptr1W116R; MetFQ28P	DutV77A
P726	MSHR6755	975	SRR6075122	MEMr (3)	—	AmrRΔV60-C63	[8]
P797	MSHR7587	437	SRR6075129	MEMr (4)	—	AmrRG30D	[8]
	MSHR7929	437	SRR6075126	SXTr (4)	AmrRG30D	BPSL2263ΔD110-G116
				MEMr (4)	—	AmrRG30D
CF11	MSHR8441	46	SRR3382162	CAZi (12)	—	PenA 10x CNV	[8,11]
				SXTr (≥32)	AmrRΔV62-H223c	Ptr1A22-G23ins_R-R-A
				Decreased DOX susceptibility (4)	AmrRΔV62-H223c	BPSL3085S130L
				Decreased MEM susceptibility (2)	—	AmrRΔV62-H223c
	MSHR8442	46	SRR3404603	SXTr (≥32)	AmrRΔV62-H223c	Ptr1A22-G23ins_R-R-A
				DOXi (8)	AmrRΔV62-H223c	BPSL3085S130L
				MEMr (3)	—	AmrRΔV62-H223c
Non-DPMS QP09	MSHR8481	1378	SRR6075123	MEMr (6)	—	AmrR∆A70-H223d	[8]
P989	MSHR9021a	132	SRR6075127	MEMr (3)		AmrRS166P; AmrRA145fsb	[8]
Strains with unusual antibiotic sensitivity
P314	MSHR1043§	131	SRR6380769	Unusual GEN sensitivity	—	AmrAL247fs	[19]
Malaysian patient	MSHR5089§	881	SRR2975737	Unusual sensitivity to aminoglycosides (GEN, KAN) and macrolides (AZM)	—	AmrBT368R	[18]
Strains with stepwise variants
P179	MSHR0535	114	SRR6075120	DOXs (4)§	BPSL3085R104fs	—	[8]
Thai patient	354e	78	AHJD00000000.1	AMCs (4/2)§	penA -78G>A	—	[9]

Abbreviations: AZM, azithromycin; CAZ, ceftazidime; CIP, ciprofloxacin; CNV, copy-number variation; DOX, doxycycline; GEN, gentamicin; KAN, kanamycin; MEM, meropenem; MIC, minimum inhibitory concentration; SXT, co-trimoxazole

^¥i, intermediate; r, resistant; s, sensitive

^†Where applicable

^‡Encodes an 800kb inversion that affects the 3’ end of bpeT, the transcriptional regulator of the BpeEF-OprC efflux pump⁹

^#False positive; isolate remained sensitive despite encoding a known AMR determinant. Cause of reversion unknown

^§Strains MSHR1043¹⁹ and MSHR5089¹⁸ encode unusual antimicrobial susceptibility, and are susceptible to all five clinically-relevant antibiotics (i.e. AMC, CAZ, DOX, MEM, SXT), and strains MSHR0535 and 354e encode stepwise variants that elevate the DOX and AMC MICs, respectively, but do not exceed the established resistance cut-off for these antibiotics

^aMSHR9021 was intentionally sequenced from a potentially mixed population to capture population diversity. Two AmrR mutants, AmrR_S166P and AmrR_A145fs, were present at ratios of ~66% and 33%, respectively⁸. In the non-mixture mode, only the dominant variant, AmrR_S166P, is detected.

^bFrameshift indel shortens protein length from 223 to 183 residues

^cFrameshift indel increases protein length from 223 to 285 residues

^dFrameshift indel shortens protein length from 223 to 117 residues

Table 2.

Burkholderia pseudomallei strains phenotypically confirmed to be sensitive towards the five clinically-relevant antibiotics, and associated genome data.

Patient ID	Isolate	ST	Genome accession	Antibiotic MIC (µg/mL)*					Reference/s
				AMC	CAZ	DOX	MEM	SXT
Australian patient	MSHR0293	236	SRR4254579	2/1	1	1	0.5	0.4	[29]
P179	MSHR0492	114	SRR6075119	1.5/0.75	1.5	1	1.5	1	[8]
	MSHR0934	114	SRR6075116	2/1	2	1	0.75	1
P337	MSHR1141	333	SRR2975732	1.5/0.75	1.5	1	0.75	0.75	[22]
P608	MSHR3763	36	SRR2887021	4/2	2	0.75	0.75	3	[8]
P726	MSHR5864	975	SRR6075121	3/1.5	1.5	1	0.75	1.5	[8]
P797	MSHR6522	437	SRR6075128	2/1	1.5	1	0.5	1.5	[8]
CF1	MSHR0913	279	SRR3404575	2/1	3	1	1	0.5	[11]
	MSHR1053	279	SRR3404578	1.5/0.75	2	1	0.5	1
Malaysian patient	MSHR5093	881	SRR2975738	6/3	4	1.5	1	3	[18]
Malaysian patient	MSHR5104	881	SRR2975740	2/1	0.75	1	0.19	0.75	[18]
CF6	MSHR5651	1040	SRR3381886	1.5/0.75	1.5	0.38	0.5	0.75	[11,25]
CF9	MSHR5662	252	SRR3381885	2	2	2	0.75	0.5	[11,25]
	MSHR5670#	252	SRR3404600	1.5/0.75	2	2	0.75	0.5
CF10	MSHR8438	442	SRR3382015	2/1	3	1	1	0.25	[11,25]
	MSHR8440	442	SRR3404601	2/1	3	1	0.75	0.38

Abbreviations: AMC, amoxicillin-clavulanate; CAZ, ceftazidime; DOX, doxycycline; MEM, meropenem; ST, multilocus sequence type; SXT, co-trimoxazole

*According to Etesting

^#False positive; isolate remained sensitive despite encoding known AMR determinants towards DOX (BPSL3085_A88fs) and SXT (Ptr1_W116R; MetF_Q28P; Dut_V77A). Cause of reversion unknown

Table 3.

ARDaP prediction in three meropenem-resistant Burkholderia pseudomallei isolates with previously unknown antimicrobial resistance (AMR) determinants.

Strain ID	ARDaP-predicted AMR determinant	MEM MIC (µg/mL)	Reference/s
MSHR1058	AmrR∆P81-H223a	12 (MEM)	[8]; This study
MSHR1174	AmrRG149fsb	6 (MEM)	This study
MSHR8777	AmrRΔA128-H223c	4 (MEM)	This study

^aPreviously undescribed 3’ amrR deletion shortens protein length from 223 to 130 residues.

^bPreviously undescribed 11bp insertion shortens protein length from 223 to 178 residues.

^cPreviously undescribed 3’ amrR insertion increases protein length from 223 to 621 residues.

Culturing, WGS, and genome assembly. B. pseudomallei culture, DNA extraction, and WGS were performed as described elsewhere [11]. Genomic data for MSHR1058, MSHR1174, and MSHR8777 were uploaded to the NCBI Sequence Read Archive database (BioProject PRJNA641249). Accession numbers for all other genomic data are listed in Tables 1 and 2. For genomes lacking a publicly-available assembly, MGAP v1.1 (https://github.com/dsarov/MGAP—Microbial-Genome-Assembler-Pipeline) was used, with archetypal strain K96243 (RefSeq accessions NC_006350.1 and NC_006351.1) provided as the scaffolding reference.

Minimum Inhibitory Concentrations (MICs). MICs were determined using Etests (bioMérieux, Murarrie, Australia). Sensitive, intermediate and resistant cut-offs were based on the Clinical and Laboratory Standards Institute (CLSI) M100-S17 guidelines for B. pseudomallei (≤8/4, 16/8, and ≥32/16 µg/mL for AMC; ≤8, 16, ≥32 µg/mL for CAZ; ≤4, 8, ≥16 µg/mL for DOX and IPM; ≤2/38, nil, ≥4/76 µg/mL for SXT). CLSI guidelines do not list MEM for B. pseudomallei; however, based on prior work [16,17], and recent proposed EUCAST breakpoints for B. pseudomallei, we categorised MEMr as ≥3 µg/mL. Likewise, the CLSI guidelines do not list gentamicin (GEN) MIC values for B. pseudomallei due to almost ubiquitous resistance (>16 µg/mL) towards this antibiotic; however, there are notable exceptions [18,19]. We chose a GEN-sensitive cut-off of ≤4 µg/mL, which also reflects those strains unable to grow on Ashdown's agar, a selective medium for B. pseudomallei isolation that contains 4 µg/mL GEN. For the four false-positive strains, Etests were performed on a minimum of two occasions by different operators to ensure MIC robustness.

AMR software parameters. The default RGI v5.1.0 database parameters of CARD v3.0.9 (https://card.mcmaster.ca/analyze/rgi; accessed 25Jun20), ARIBA v2.14.5 (https://github.com/sanger-pathogens/ariba), ResFinder v4.1 (https://cge.cbs.dtu.dk/services/ResFinder/), and AMRFinderPlus v3.8.28 (https://www.ncbi.nlm.nih.gov/pathogens/antimicrobial-resistance/AMRFinder/) were examined for performance across the B. pseudomallei genomes.

ARDaP AMR database construction. ARDaP is available at: https://github.com/dsarov/ARDaP. All reported B. pseudomallei AMR determinants, including stepwise AMR mutations and unusual antimicrobial susceptibility mutations (Table 1; Table S1), were annotated relative to K96243. The AMR determinants (as of version 1.7) are summarised in an SQLite database (Table S1; most up-to-date version available at: https://github.com/dsarov/ARDaP/tree/master/Databases/Burkholderia_pseudomallei_k96243). Briefly, CAZ resistance (CAZr) is caused by altered PenA β-lactamase substrate specificity [20], [21], [22], [23], penA upregulation [9,22,24,25] (including CNVs [11]), or loss of penicillin-binding protein 3 [26]; AMC resistance (AMCr) is caused by penA upregulation [19,22]; MEMr is caused by AmrAB-OprA, BpeAB-OprB, or BpeEF-OprC resistance-nodulation-division (RND) multidrug efflux pump regulator loss-of-function [8]; SXT intermediate (SXTi) or full resistance (SXTr) is caused by cumulative mutations in core metabolism pathways coupled with AmrAB-OprA, BpeAB-OprB, or BpeEF-OprC RND efflux pump regulator loss-of-function [8,11,27,28]; and DOX intermediate (DOXi) or full resistance (DOXr) is caused by loss-of-function mutations within the SAM-dependent methyltransferase gene, BPSL3085, often in combination with AmrAB-OprA, BpeAB-OprB, or BpeEF-OprC regulator loss-of-function [29]. Our B. pseudomallei ARDaP database also includes AmrA and AmrB mutants that are associated with unusual aminoglycoside and macrolide susceptibility [18,19]. To avoid poor-quality WGS data or incorrect species assignments, the database also includes two conserved genetic targets (Table S1) found only in this bacterium; strains lacking these loci are flagged for further user assessment.

ARDaP algorithm. To achieve high-quality variant calls, ARDaP incorporates several tools into its workflow (full list available at: https://github.com/dsarov/ARDaP). In addition to an organism-specific SQLite database (https://github.com/dsarov/ARDaP/tree/master/Databases), ARDaP requires WGS data, either genomes or metagenomes in paired-end Illumina v1.8+ FASTQ format, or assembled genomes in FASTA format, as input (Fig. 1). For genomes in FASTA format, ARDaP first converts to synthetic Illumina v1.8+ reads using ART (version Mount Rainier 2016-06-05).[30] For genomes in FASTQ format, ARDaP performs quality filtering using Trimmomatic v0.39 followed by optional random down-sampling to a user-defined coverage (default=50x) using Seqtk (https://github.com/lh3/seqtk) to permit more rapid analysis. ARDaP then performs comparative genomic analysis to identify AMR determinants by mapping reads against an annotated reference using BWA-MEM (v0.7.17-r1188),[31] followed by SAMTools (v1.9)[32] for alignment processing and BAM creation, Genome Analysis Toolkit (GATK v4.1.0.0)[33] for SNP and indel identification, Mosdepth (v0.2.3)[34] for coverage assessment, Pindel (v0.2.5b9)[35] for CNV detection, and DELLY (v0.8.3)[36] for inversion identification. High-quality genetic variants (SNPs, indels [<50bp], CNVs, gene gain, inversions, and gene loss or truncation) are then annotated with SnpEff (v4.3.1t).[37] ARDaP next interrogates two databases: i) a customisable CARD[5] database is screened to identify horizontally-acquired AMR genes and to ignore conserved genes that do not confer AMR, and ii) a bespoke AMR determinant database (in this study, a B. pseudomallei database) containing species-specific AMR determinants. ARDaP databases are created in SQLite and can be readily updated as additional AMR determinants are identified. This database also accommodates stepwise mutations and AMR conferred by ≥2 mutations. Finally, ARDaP can predict AMR by identifying novel high-consequence mutations (i.e. those resulting in a frameshift or nonsense mutations, loss of coverage, or inversion) in known AMR genes. These putative mutants can then be flagged for further investigation with phenotypic AMR testing. ARDaP outputs are presented in a comprehensive, human-readable report (Fig. 3).

Fig. 1.

ARDaP pipeline. The user inputs assembled genome/s or raw sequencing reads. ARDaP then performs read alignment, read processing, deduplication, and variant identification. An optional phylogenetic analysis is also performed (if specified). Coverage assessment is undertaken on either single or mixed genomes (if specified); genetic variants are then annotated and antimicrobial resistance database/s interrogated. Finally, ARDaP produces a summary report of antimicrobial resistance determinants for each strain (Figure 2). *Downsampling is carried out by default but can be turned off using the --size 0 flag in ARDaP.

Fig. 3.

Operon organisation of the Burkholderia pseudomallei AmrAB-OprA resistance-nodulation-division efflux pump and loss-of-function mutations in its TetR-type regulator, AmrR. A. Transcriptional organisation of the amrR (BPSL1805), amrA (BPSL1804), amrB (BPSL1803) and oprA (BPSL1802) operon, and summary of how (i) amrR mutations cause (ii) loss-of-function of AmrR, which (iii) no longer represses expression of the resistance-nodulation-division AmrAB-OprA efflux pump, resulting in (iv) efflux pump over-expression and resistance to meropenem and aminoglycoside antibiotics. B. Distribution and annotation of amrR mutations. Eleven previously observed amrR mutations (in black)[8] have been augmented with three novel mutations identified in the current study (orange); AmrR_G149fs, AmrR_∆P81-H223, and AmrR_∆A128-H223, all of which cause amrR loss-of-function, resulting in efflux pump overexpression and meropenem resistance.

Mixture detection. ARDaP incorporates a minor allelic variant analysis function to permit variant identification from mixed genomes/metagenomes, enabling the detection of emerging AMR determinants (down to 5% abundance). Minor-variant SNPs and indels are identified using the ploidy-aware HaplotypeCaller tool in GATK v4.1; deletions and CNVs are identified with the ploidy-aware function of Pindel. B. pseudomallei strains with known AMR status were mixed at ratios of 5% increments ranging from 5:95 to 95:5, to 55-60x total depth. Two mixtures were created: MSHR0913 (sensitive to all clinically-relevant antibiotics) and MSHR5654 (SXTr, CAZr, ciprofloxacin-resistant), and MSHR0913 and MSHR8441 (SXTr; intermediate resistance to CAZ; decreased susceptibility to MEM and DOX). MSHR5654 and MSHR8441 were chosen as they represent a wide spectrum of clinically-relevant AMR and mutation types (Table 4).

Table 4.

ARDaP detection limits for AMR determinants in synthetic mixtures.

AMR determinant	K96243 gene ID	Mutation type	AMR minor allele detection (%)
Mixture #1: MSHR0913 (sensitive strain) and MSHR5654 (resistant to CAZ, CIP, and SXT)
BpeTT314fs	BPSS0290	Gene loss/truncation	10
Ptr1R21fs	BPSS0039	Gene loss/truncation	10
penA 30x	BPSS0946	Copy-number variation (30x)	5
PenAC75Y	BPSS0946	Missense SNP	5
GyrAY77S	BPSL2521	Missense SNP	10
Mixture #2: MSHR0913 (sensitive strain) and MSHR8441 (resistant to SXT; intermediate resistance to CAZ; decreased susceptibility to MEM and DOX)
AmrRΔV62-H223	BPSL1805	Gene loss/truncation	50†
BPSL3085S130L	BPSL3085	Missense SNP	15
Ptr1A22-G23ins_R-R-A	BPSS0039	In-frame insertion	10
penA 10x	BPSS0946	Copy-number variation (10x duplication)	5

AMR, antimicrobial resistance; CAZ, ceftazidime; CIP, ciprofloxacin; DOX, doxycycline; MEM, meropenem; ND, not detected; SNP, single-nucleotide polymorphism; SXT, co-trimoxazole

^†Represents the lowest allele frequency that was consistently detected in our mixed strain dataset.

3. Role of funders

This study was funded by the National Health and Medical Research Council (awards 1046812, 1098337, and 1131932 [the HOT NORTH initiative]). DEM was supported by an Australian Government Research Training Scholarship. ES was supported by an International Postgraduate Research Scholarship from James Cook University. EPP and DSS were supported by Advance Queensland fellowships (awards AQIRF0362018 and AQRF13016-17RD2, respectively). The funders had no role in study design; in the collection, analysis, or interpretation of data; in the writing of this report; or in the decision to submit for publication. The corresponding author had full access to all the data in the study and has final responsibility for the decision to submit for publication.

4. Results

Performance of existing AMR tools in B. pseudomallei. The validated dataset of 47 B. pseudomallei isolates was used to assess the performance and capacity of existing AMR tools to identify AMR determinants in AMR but not antimicrobial-sensitive strains. According to CARD and AMRFinderPlus, all 47 genomes were found to harbour AMR determinants; however, these determinants corresponded with conserved genes in all B. pseudomallei (Table S2). In addition, CARD, ResFinder, and AMRFinderPlus failed to identify any clinically relevant AMR determinants in the 25 AMR strains. ARIBA outperformed CARD and AMRFinderPlus due to its ability to include missense – although not nonsense – SNPs in its database construction, and to identify SNPs and indels in its report outputs. However, ARIBA cannot identify CNVs or inversions, and it requires considerable user expertise and assessment time to determine the validity of variant outputs and to distinguish real AMR determinants from natural variation. Due to these limitations, we did not pursue this tool further.

ARDaP development and performance in B. pseudomallei. Given the shortcomings of existing AMR software, ARDaP was designed to both identify known AMR determinants and to ignore non-causal genetic variants (Table 2). When tested against the 47 validated isolates, ARDaP correctly identified all B. pseudomallei AMR determinants (Table 1) and yielded no false negatives; however, four false positives (MSHR5654, MSHR5666, MSHR5669, and MSHR5670), all of which were isolated from chronic cystic fibrosis (CF) infections, were identified. The first of these, MSHR5654 (from CF6) [11], was predicted to be MEMr due to the presence of BpeT_Thr314fs. The remaining false-positive strains (all from CF9) [11], encode a SAM-dependent methyltransferase truncation (BPSL3085_A88fs) [25] and were predicted by ARDaP to be DOXr. We also observed BPSL3085_A88fs in an unrelated DOXr chronic CF strain, Bp1651 [38] (Table 1). BPSL3085 mutations confer DOXr likely by altering ribosomal methylation patterns [11,29]. However, all three strains remained DOX-sensitive (1.5 µg/mL) despite other CF9 strains encoding BPSL3085_A88fs and being DOXr (MSHR5665: MIC=6 µg/mL; MSHR5667: MIC=48 µg/mL; Table 1) [11]. The much higher DOX MIC in MSHR5667 is attributable to a second mutation (AmrR_L132P; Table 1).

Importance of including natural genetic variation in AMR databases. Accurate prediction of novel AMR determinants requires thorough cataloguing of both confirmed AMR-causing mutations and natural variation in AMR-encoding genes to avoid false positives. To illustrate this point, a PenA β-lactamase missense mutation (K96243 numbering: PenA_S78F; encoded by BPSS0946) has previously been linked to AMCr [20,23,38]. However, we found that PenA_S78F alone is unlikely to cause AMCr due to its presence in genetically diverse AMC-sensitive strains (MIC=3-4 µg/mL in strains MSHR0291, MSHR0668, MSHR0848, MSHR0911, MSHR1711, MSHR2212, MSHR3902, MSHR4797, MSHR8392, and MSHR9887). Instead, AMCr is likely conferred by both PenA_S78F and penA upregulation, the latter of which can be caused either by mutations within the 5’ untranslated region [22], or by penA CNVs [11]. We therefore included PenA_S78F as a putative stepwise mutation in the B. pseudomallei ARDaP database (Table 2), with an additional penA upregulation mutation required to confer the AMCr phenotype. In another example, we observed that both AMR and antimicrobial-sensitive strains can possess 3’-truncated amrR (Table 2). Multiple frameshift mutations and deletions in amrR are associated with MEMr[8] due to loss-of-repressor function (Fig. 3). However, certain 3’ region mutations (i.e. those affecting residues ~210-223) do not cause MEMr (Table 2). To accommodate this natural genetic variation, we coded ARDaP to ignore these non-causal 3’ variants, thereby greatly reducing false-positive MEMr rates.

AMR predictive capacity of ARDaP. We tested ARDaP's predictive capacity to identify the causative mutation/s in three clinical MEMr strains (MSHR1058, MSHR1174, and MSHR8777; MEM MIC range: 4–12 µg/mL; Table 3) with no previously reported AMR determinants. All patients had received MEM treatment prior to isolate retrieval. ARDaP identified novel amrR mutations in each strain, all of which resulted in AmrR loss-of-function (AmrR_∆P81-H223 in MSHR1058; AmrR_G149fs in MSHR1174; AmrR_ΔA128-H223 in MSHR8777; Fig. 3).

Reversions and unusual antimicrobial susceptibility. Aminoglycoside- and macrolide-class antibiotics are typically not included in melioidosis treatment regimens due to near-ubiquitous intrinsic resistance; indeed, GEN resistance is commonly used for B. pseudomallei selection [18]. However, rare cases of sensitivity have been documented, such as in ST-881 and ST-997 strains from Sarawak, Malaysian Borneo, which naturally encode AmrB_T368R, resulting in AmrAB-OprA loss-of-function and unusual aminoglycoside and macrolide susceptibility [18]. An AmrAB-OprA loss-of-function variant has also been described in Bp1651 (AmrB_A254fs) [38]. Although ARDaP detected amrR loss in MSHR1043, the co-presence of amrA loss (AmrA_L247fs) resulted in reversion of MEMr to a wild-type MIC (0.75 µg/mL). This reversion also causes unusual gentamicin (MIC=1 µg/mL) [19] and presumably kanamycin and azithromycin sensitivity. Given their confounding potential, we incorporated these reversions into ARDaP to more accurately reflect the true strain phenotype.

ARDaP performance on mixed sequence data. To assess the performance of the mixture function in ARDaP, Illumina reads from antimicrobial-sensitive and AMR B. pseudomallei strains (Table 4) were mixed at ratios ranging from 5:95 to 95:5. ARDaP identified three AMR determinants down to the lowest tested ratio of 5% minor allele frequency: a penA 10x CNV from MSHR8441, a penA 30x CNV in MSHR5654, and PenA_C69Y (K96243 numbering: PenA_C75Y) in MSHR5654 (Table 4). The other determinants were identified by ARDaP when present at minor allele frequencies of 10% (Ptr1_R21fs, Ptr1_{A22_G23ins_R-R-A}, BpeT_T314fs, and GyrA_Y77S), 15% (BPSL3085_S130L), and 50% (AmrR_ΔV62-H223) (Table 4). The high sensitivity of PenA_C75Y detection can be explained by the multicopy nature of this gene in MSHR5654; this missense variant likely has a sensitivity closer to 10–15% when present as a single copy, as observed with missense variants Gyr_Y77S and BPSL3085_S130L (Table 4). Gene truncations (e.g. AmrR_ΔV62-H223) had the lowest sensitivity.

Next, ARDaP was tested on a previously detected AmrR mixture from strain MSHR9021, which encodes AmrR_S166P and AmrR_A145fs variants at ~66% and ~33% allele frequencies, respectively [8] (Table 1). ARDaP detected AmrR_S166P and AmrR_A145fs at allele frequencies of 63% and 31%, respectively, thus closely reflecting their known proportions.

ARDaP reports. ARDaP generates an easy-to-interpret report that summarises the AMR determinants and associated antibiotic phenotype/s for each genome (Fig. 2). This report summarises AMR findings for first-line, second-line, and tertiary antibiotics, along with instances of unusual antibiotic susceptibility, and has been designed to prioritise a clinical workflow. In addition, the ARDaP report lists stepwise AMR determinants, thereby informing early treatment shifts aimed at mitigating the risk of AMR emergence and fixation.

Fig. 2.

Example antimicrobial resistance (AMR) summary report produced by ARDaP for strain MSHR5654. The final step in the ARDaP pipeline is the production of a user-friendly report that summarises patient and sample details, confirms that the given isolate was interrogated against the correct database (in this case, Burkholderia pseudomallei), identifies any predicted AMR (including what mutation/s has/have been detected) and what antibiotic/s have been affected, identifies unusual antimicrobial sensitivity and natural variation that does not confer AMR, and identifies stepwise mutations that lower the barrier to AMR development.

5. Discussion

This study describes the development and first-described implementation of the new AMR tool, ARDaP, for truly comprehensive AMR determinant identification from NGS and genome assembly data, including from mixed (e.g. metagenomic) data. Using the melioidosis pathogen, B. pseudomallei, as a model organism, we demonstrate that ARDaP provides several key advantages over existing AMR software. From 47 well-validated isolate genomes, we found that CARD, ResFinder, and AMRFinderPlus failed to identify any AMR determinants in B. pseudomallei, and for CARD and AMRFinderPlus, all 47 isolates were found to harbour AMR determinants, despite 16 being antimicrobial-sensitive. Two reasons underpin this shortcoming of existing AMR software: first, B. pseudomallei exclusively acquires AMR through chromosomal mutation, thereby limiting the value of tools that are heavily biased towards gene gain identification; and second, these AMR tools are unable to identify AMR variants conferred by indels, CNVs, inversions, or gene loss/truncation. Although ARIBA can detect indels, in our hands, this tool provided comparable information to variant report outputs generated by comparative genomic pipelines; the user requires extensive domain-specific knowledge to accurately identify and interpret outputs, particularly when trying to differentiate AMR-conferring variants from naturally-occurring genetic variation. Other shortcomings of ARIBA include cumbersome and labour-intensive input file requirements, restrictions on database construction (e.g. reference genomes with indels and nonsense mutations cannot be included), and an inability to identify CNVs. In contrast, ARDaP uses standardised variant annotation, can differentiate natural gene variation from known and putative AMR determinants, can detect CNVs and inversions, and provides a user-friendly output that does not require domain-specific knowledge for accurate interpretation.

Assessment of ARDaP's performance across the 47 characterised isolates demonstrated that this software accurately identified all AMR determinants (including stepwise variants) in all strains, except for four false positives. The first of these, MSHR5654, was predicted to be MEMr due to a BpeT truncation;[11] however, Etesting showed MEM sensitivity (2 µg/mL) in this strain, just below the MEMr threshold (≥3 µg/mL) [11,25]. Although alterations in BpeT have been putatively linked with MEMr in MSHR1300 (4 µg/mL) [8] and 354e (6 µg/mL),[9] the role of BpeT mutations in conferring MEMr is contentious[27]. In support of this notion, MSHR1300 also encodes AmrR_K13fs, a TetR-family cis-acting repressor of the AmrAB-OprA RND efflux pump, which likely causes MEMr in its own right[8], and in 354e, the ~800kb inversion likely also affects other AMR-conferring genes besides bpeT. As such, the B. pseudomallei ARDaP database was updated to flag bpeT variants as stepwise mutations rather than solely conferring MEMr (Table 1; Table 2), thereby correcting the original false-positive call for MSHR5654. This issue highlights the complexity of unravelling AMR determinants, and in this case, the need for additional work to determine a role, if any, for bpeT mutations in conferring MEMr.

The remaining three false-positive strains, all of which are longitudinal isolates retrieved from a single chronic CF airway infection (patient CF9) [11], were predicted by ARDaP to be DOXi or DOXr due to the presence of a BPSL3085_A88fs variant in these strains [25]. Indeed, other CF9 isolates that harbour the BPSL3085_A88fs variant exhibit DOXi (MSHR5665; MIC=6) or DOXr (MSHR5667; MIC=48) phenotypes (Table 1). This variant is also found in unrelated strains MSHR3683 (DOXi) and Bp1651 (DOXr). Taken together, there is strong evidence that BPSL3085_A88fs confers DOXi or DOXr in B. pseudomallei. We thus postulate that MSHR5666, MSHR5669, and MSHR5670 encode an unidentified mutation that reverts them to a DOX-sensitive phenotype. Notably, all longitudinal CF9 isolates, including MSHR5666 and MSHR5669, encode mutS mutations, resulting in a hypermutator phenotype [11,25]. Therefore, identifying the causal basis for this reversion is non-trivial due to the large number of mutations (range: 112–157) accrued by these hypermutator strains [11]. In addition, MSHR5670 was predicted to be SXTr due to Ptr1_W116R, MetF_Q28P, and Dut_V77A variants, yet exhibited SXT sensitivity (Table 2). The cause of SXTr reversion in this strain is also currently unknown and requires further exploration. Our results show that chronic infections, particularly those in which hypermutated strains have emerged, represent the most challenging scenario from which to accurately predict AMR phenotypes. We therefore recommend that chronically infecting strains be subjected to conventional phenotypic testing to confirm AMR profiles predicted from NGS data.

In most melioidosis treatment guidelines, IPM has been replaced by MEM due to neurotoxicity concerns [39]. However, the recent discovery of MEMr B. pseudomallei has resurrected IPM as a treatment option due to a lack of cross-resistance between these carbapenems[8] and exceedingly low rates of reported IPM resistance (IPMr) [40]. The one study reporting an IPMr (MIC=8 µg/mL) B. pseudomallei strain, Bp1651, attributed this phenotype to a PenA_T147A mutation (K96243 numbering: PenA_T153A) combined with penA upregulation due to a promoter mutation [38]. We subsequently refuted the role of the PenA_T147A variant alone in conferring IPMr by identifying three genetically unrelated PenA_T153A-encoding strains that were IPM-sensitive [8]. Further, this variant is dominant (>50%) in publicly available B. pseudomallei genomes, none of which have been reported as IPMr. Given that PenA_T147A occurs at a very high rate in the wild-type B. pseudomallei population, and none have been shown to exhibit IPMr, this mutant has not been included in our ARDaP database. However, this variant can readily be added as a stepwise AMR determinant should further evidence come to light about its role in conferring AMR.

ARDaP has not just been designed to detect known AMR determinants; its databases can also be configured to ignore natural variation, and to predict novel AMR variants from known AMR loci, both of which are essential facets of accurate AMR prediction from WGS data. For example, our initial analyses identified several amrR mutants that were predicted to confer MEMr. However, Etesting of these strains showed that most strains were MEM-sensitive. Closer inspection of the amrR gene found considerable variability spanning residues 210 to 223 in these strains, indicating that these 3’ mutations do not impact the regulator or repressor activity of the AmrAB-OprA RND efflux pump. By ignoring this highly mutable portion of the amrR gene, we dramatically reduced the number of false positive AMR determinants identified by ARDaP.

To predict AMR determinants, ARDaP will flag known AMR genes encoding novel high-consequence (i.e. nonsense, frameshift, or gene loss) mutations for further user assessment. This prediction capacity of ARDaP was tested in three previously genetically uncharacterised MEMr strains: MSHR1058, MSHR1174, and MSHR8777, each of which was isolated from clinical infections where MEMr emerged during MEM therapy. In each case, ARDaP identified novel, high-consequence mutations affecting amrR, the local regulator of the AmrAB-OprA RND efflux pump (Fig. 3; Table 3). This result provides further confirmation of the link between MEM administration and potential treatment failure due to AmrR mutability [8], and demonstrates the value of ARDaP for predicting novel AMR determinants.

Genetic variants conferring unusual antimicrobial susceptibility, including those brought about by reversions, represent an important yet commonly overlooked aspect of AMR detection and prediction software. Most B. pseudomallei strains are naturally resistant to aminoglycosides and macrolides, meaning that these antibiotic classes are almost universally excluded from melioidosis treatment regimens due to inherent AMR towards these antibiotic classes; however, there are notable exceptions. For example, certain B. pseudomallei clones from Malaysian Borneo are naturally susceptible to gentamicin, kanamycin, and azithromycin due to AmrAB-OprA loss-of-function [18], and this phenotype can also arise in vivo due to within-host evolution. Importantly, such strains can conceivably be effectively treated with aminoglycoside and macrolide antibiotics, which are not typically considered for melioidosis treatment due to assumed inherent resistance. Strains encoding AmrAB-OprA loss-of-function variants (e.g. AmrB_T368R, AmrB_A254fs, AmrA_L247fs) are also at far lower risk of developing MEMr than wild-type strains due to the abrogation of deleterious amrR mutations that would otherwise cause MEMr. The identification of strains encoding amrAB-oprA loss-of-function mutations would thus strongly support long-term MEM use due to a far lower risk of MEMr development in such cases. These findings highlight the value of including sensitivity-conferring variants in AMR databases by increasing the antibiotic arsenal in naturally multidrug-resistant pathogens where treatment options are limited.

The ARDaP algorithm is mixture-aware, an important feature for detecting emerging AMR determinants in mixed strain data (e.g. non-purified colonies, culture sweeps, total clinical specimens). Using mixtures of AMR and antimicrobial-sensitive strains at varying ratios, we defined the limits of mixture detection in ARDaP for common AMR variants in B. pseudomallei. Overall, ARDaP confidently identified AMR determinants in the tested mixtures, albeit with varying sensitivities. CNVs were most readily detected by ARDaP, with 10x and 30x CNVs able to be distinguished at the lowest tested allele frequency of 5%. AMR-conferring SNPs and indels were robustly detected at minor allele frequencies of 10-15% (Table 4). Gene truncations were the least sensitive AMR variant type to detect from mixtures, with the one truncation examined in this study (AmrR_ΔV62-H223) only detectable when present at ≥50% allele frequency. A possible explanation for the much lower sensitivity of gene truncation variant detection in mixed data is the challenge of discriminating gene loss from Illumina depth coverage variation, coupled with inherent limitations in short-read data mapping. Further validation of specific variant mixtures is recommended when new mixtures are identified to determine their sensitivity. In addition, deeper sequencing (e.g. 100–500x) should enable more robust mixture detection at lower allele frequencies.

The easy-to-interpret AMR summary report generated by ARDaP (Fig. 2) represents a major improvement over current AMR software such as AMRFinderPlus, ARIBA, CARD, and ResFinder, which require an intimate understanding of AMR determinants to correctly interpret outputs and to ignore naturally occurring genetic variation. The AMR report produced by ARDaP represents a crucial step towards the incorporation of WGS as a routine tool for guiding best-practice AMR stewardship and personalised treatment regimens in the clinical diagnostic setting, and will help to accelerate the translation of NGS-to-bedside diagnostics.

Caveats and Limitations. We acknowledge that there are several limitations to our study. First, we have, to date, only developed one pathogen-specific AMR database for ARDaP; additional databases need to be populated for other microbes of interest, the curation of which is time-consuming and laborious. To begin addressing this task, we are currently developing ARDaP-compatible AMR databases for Haemophilus influenzae and Pseudomonas aeruginosa. Second, B. pseudomallei is hyperendemic in many resource-poor tropical regions, where access to NGS platforms and bioinformatics expertise is limited or non-existent. Therefore, ARDaP is unlikely to guide public health interventions in these regions until NGS capacity is better developed and funded, meaning that a large proportion of AMR infections in melioidosis-endemic regions will remain undetected. Despite this shortcoming, ARDaP provides a major advance towards the routine use of NNNGS for rapid and accurate acquired AMR detection in B. pseudomallei in well-resourced settings, and will be essential for informing treatment shifts and improving patient outcomes. Third, the lack of B. pseudomallei human-to-human and zoonotic transmission limits the use of AMR prediction in B. pseudomallei to individual cases rather than for larger epidemiological studies (e.g. outbreak tracking or global AMR dispersal). Finally, our study only included 25 AMR strains, the majority of which have been identified from our isolate collection. Whilst modest, these strains represent all publicly available, global, nonredundant AMR B. pseudomallei strains. Dual-use concerns in Select Agent pathogens such as B. pseudomallei mean that it is not possible to induce AMR in the laboratory setting, which has hampered the identification of novel AMR determinants as AMR identification is only possible from infected hosts. More work is needed to identify AMR strains and their associated determinants in B. pseudomallei, particularly from melioidosis hotspots in Asia, Africa, and Central and South America.