Optimized Method for Bacterial Nucleic Acid Extraction from Positive Blood Culture Broth for Whole-Genome Sequencing, Resistance Phenotype Prediction, and Downstream Molecular Applications

ABSTRACT

The application of direct metagenomic sequencing from positive blood culture broth may solve the challenges of sequencing from low-bacterial-load blood samples in patients with sepsis. Forty prospectively collected blood culture broth samples growing Gram-negative bacteria were extracted using commercially available kits to achieve high-quality DNA. Species identification via metagenomic sequencing and susceptibility prediction via a machine-learning algorithm (AREScloud) were compared to conventional methods and other rapid diagnostic platforms (Accelerate Pheno and blood culture identification [BCID] panel). A two-kit method (using MolYsis Basic and Qiagen DNeasy UltraClean kits) resulted in optimal extractions. Taxonomic profiling by direct metagenomic sequencing matched conventional identification in 38/40 (95%) samples. In two polymicrobial samples, a second organism was missed by sequencing. Prediction models were able to accurately infer susceptibility profiles for 6 common pathogens against 17 antibiotics, with an overall categorical agreement (CA) of 95% (increasing to >95% for 5/6 of the most common pathogens, if Klebsiella oxytoca was excluded). The performance of whole-genome sequencing (WGS)–antimicrobial susceptibility testing (AST) was suboptimal for uncommon pathogens (e.g., Elizabethkingia) and some β-lactamase inhibitor antibiotics (e.g., ticarcillin-clavulanate). The time to pathogen identification was the fastest with BCID (1 h from blood culture positivity). Accelerate Pheno provided a susceptibility result in approximately 8 h. Illumina-based direct sequencing methods provided results in time frames similar to those of conventional culture-based methods. Direct metagenomic sequencing from blood cultures for pathogen detection and susceptibility prediction is feasible. Additional work is required to optimize algorithms for uncommon species and complex resistance genotypes as well as to streamline methods to provide more rapid results.

INTRODUCTION

Bloodstream infections are a major cause of morbidity and mortality, and rapid pathogen identification and antimicrobial susceptibility phenotyping are critical to patient outcomes (1, 2). Currently, pathogen identification and phenotypic antimicrobial susceptibility testing can take up to 3 days or longer. Consequently, rapid molecular detection and gene profiling methodologies are desirable (2–4). There could be a great benefit in a feasible holistic approach to attain single-host-depleted DNA extraction directly from blood culture (BC) broth, suitable for multiple downstream molecular applications. Furthermore, the implementation of commercial kits with minimal out-of-kit modifications and potential automation would ensure robust, reproducible results.

We aimed to develop a reliable DNA extraction method for direct whole-genome sequencing (WGS) from positive blood culture broth to support rapid bacterial characterization and antibiotic susceptibility prediction in patients with bloodstream infections. The primary aim was to obtain human-depleted and enriched microbial DNA extracts from blood culture broth suitable for multiple downstream molecular applications, including WGS. The objective was to achieve high-quality bacterial DNA extracts, depleted of human DNA and inhibitors (salts, proteins, enzymes, preservatives, and neutralizing compounds), of appropriate input lengths and at usable concentrations for WGS. Additionally, the effects of benchtop durations or freeze-thaw conditions were assessed. Downstream applications included traditional and real-time PCR as well as short- and long-read WGS using either enzyme-based or ligation library preparations. As the ultimate goal of molecular testing from positive blood culture broth is clinical implementation (5), the results obtained were assessed for their ability to predict clinically relevant microbial phenotypes (e.g., species identification, resistance gene detection, and antibiotic susceptibility testing [AST]). To evaluate the ability to rapidly predict AST results in silico directly from genomic data, we compared a machine-learning (ML)-based WGS-AST tool to conventional culture-based methods. In addition, we compared these methods to commercially available rapid diagnostic systems based on morphokinetic cellular analysis and multiplex PCR.

MATERIALS AND METHODS

Blood cultures.

Forty blood culture broth samples (FA plus, FN plus, and pediatric PF plus bottles; bioMérieux) that were flagged as positive for mono- or polymicrobial growth with Gram-negative bacteria (identified by Gram staining and microscopy) were included. These samples were collected from patients presenting to emergency departments or admitted to an intensive care unit (ICU) served by the Central Laboratory, Pathology Queensland, Brisbane, Australia. Two clinical samples containing CTX-M extended-spectrum-β-lactamase (ESBL)-producing Escherichia coli (DETECT-110 and DETECT-111) and one sample containing non-ESBL-producing E. coli (DETECT-112) were used for assay development and validation purposes. Two samples (DETECT-113 and DETECT-114) had no positive growth and were used as negative controls. The 37 subsequent samples (DETECT-115 to DETECT-151) were tested prospectively across all testing platforms. Thus, a total of 40 samples were included in the metagenomic analysis, but only 37 were tested across all diagnostic modalities (Fig. 1). Positive blood culture broth samples were deidentified and analyzed at the University of Queensland (UQ) Centre for Clinical Research. Blood culture bottles were removed from the BacT/Alert Virtuo system (bioMérieux) once they were flagged as positive and extracted by the research laboratory within 1.5 h.

FIG 1.

Study workflow. Direct metagenomic sequencing from positive blood culture broth samples was compared with culture-based methods and commercial rapid diagnostics. Two DNA extraction methods were compared, each using half (n = 20) of the samples. AxDx, Accelerate Pheno; BCID, blood culture identification PCR panel (bioMérieux); gDNA, genomic DNA; SPRI, solid-phase reversible immobilization; ID, identification; AST, antimicrobial susceptibility testing; WGS, whole-genome sequencing; MALDI-TOF, matrix-assisted laser desorption ionization–time of flight; AMR, antimicrobial resistance; ML, machine learning; ESBL, extended-spectrum β-lactamase; p-AmpC, plasmid-mediated AmpC β-lactamase.

Sample processing and storage.

Genomic DNA (gDNA) was extracted from blood culture broth samples upon receipt, and the remaining samples were frozen at −20°C and −80°C and, if required, thawed to room temperature from frozen.

Genomic DNA extraction.

Host gDNA was depleted using the MolYsis Complete and MolYsis Basic kit 0.2-mL and 1-mL protocols (Molzyme, Germany), according to the manufacturer’s instructions, with the following exceptions. The blood culture broth starting volume for the Basic kit was increased from 0.2 to 0.5 mL for higher yields, and the manufacturer’s instructions were followed (see the supplemental material for additional details). After the host depletion stage, samples were centrifuged at 10,000 × g for 30 s, and the supernatant was removed. The microbial pellet underwent gDNA extraction by one of two methods (Mini-Pure extraction [method 1] and UltraClean extraction [method 2] [for full details, see the supplemental material]). Both methods were shown to produce adequate DNA for sequencing and other downstream applications, yet the UltraClean method required a reduced extraction time (Fig. 2) and fewer steps in the workflow, so this method was therefore preferred.

FIG 2.

Comparison of times to results for conventional culture, metagenomic WGS from positive blood culture broth (for short reads), and commercial rapid diagnostic platforms.

Genomic DNA quality and purity checks were undertaken using a Qubit fluorometer (Life Technologies), a NanoDrop 2000 spectrophotometer (Thermo Scientific), and an Agilent TapeStation 4150 instrument using genomic DNA ScreenTape and reagents or the D1000 high-sensitivity system for sequencing library preparations. In addition, we assessed the utility of gDNA size selection using the Circulomics short-read eliminator, inhibitor removal by the Qiagen DNeasy PowerClean Pro cleanup kit, and the NEBNext microbiome DNA enrichment kit ethanol precipitation protocol modified by the removal of Tris-EDTA (TE) buffer and 12.5 μL of prewarmed water for elution. Twenty Mini-Pure extracts (method 1) and 20 UltraClean extracts (method 2) were prepared for short-read sequencing (Fig. 1).

Inhibitor removal and human genomic DNA depletion.

Traditional and real-time PCR assays were undertaken to determine any effects of sodium polyanetholesulfonate (SPS) being copurified or acting as an inhibitor of PCR. Inhibition was assessed according to the positive controls and experimental outcomes outlined previously by Regan et al. (6). Dilutions were prepared from 1:10 to 1:10,000. There was no inhibition of PCR amplification as there were resolved bands of the expected amplicon sizes for each dilution. This was observed with real-time PCR with cycle threshold (C_T) values of 13 at 1:10, with 3-fold increases to a C_T of 27 for 1:10,000 dilutions (see Fig. S1 in the supplemental material). TaqMan endogenous retrovirus-3 gene (ERV-3) real-time PCR was undertaken to determine if host human genomic DNA was depleted in the subsequent microbial DNA extraction step. The assay confirmed complete human genomic DNA depletion (Fig. S2A) and 16S gene detection for microbial DNA extraction (Fig. S2B).

Blood culture extraction under various storage temperatures and durations.

While all prospectively collected samples underwent DNA extraction upon receipt, the viability of storage conditions for blood culture broth samples was also assessed for concentration, purity, and length under two different conditions: room temperature for up to 5 days and thawed from frozen. DNA was extracted directly from BCs by using a MolYsis kit and a minikit without the solid-phase reversible immobilization (SPRI) bead cleanup step on day 1 and every 24 h until day 5. The same BC sample was frozen at −20°C on day 1, thawed to room temperature on day 5, and extracted as described above. Overall, the DNA extraction concentration, purity, and length with storage at room temperature for 2 to 5 days and frozen were comparable to those of the fresh BC baseline extraction (day 1).

Sequencing.

Short-read sequencing from blood culture broth extracts was performed by utilizing the Nextera DNA Flex library prep kit (Illumina), with a modification of the starting input of 5 μL gDNA for samples of >20 ng/μL. Pooled libraries were loaded into a mid- or high-output reagent cartridge (300 cycles) and sequenced as paired-end reads on the Illumina MiniSeq platform. Long-read sequencing was undertaken with R9.4.1 flow cells in a singleplex format using either the library preparation rapid sequencing or ligation kits and in a multiplex format using the rapid barcoding kit or ligation kit with native barcodes (Oxford Nanopore Technologies). All sequencing was performed by utilizing a flow cell priming kit (catalog number EXP-FLP002), and voltage drift was accounted for where the flow cell went through a wash protocol.

Metagenome assemblies, taxonomic profiling, and WGS-AST using machine learning.

Assembly and binning for whole-genome sequencing–antimicrobial susceptibility testing (WGS-AST) from metagenomes were performed using the nf-core/mag v2.1.1 pipeline (7). In short, raw reads were trimmed and mapped against the GRCh38 and PhiX genomes to remove reads from contaminant species. The retained reads were assembled using both SPAdes (8) and MEGAHIT (9). The binning of assembled metagenomes into metagenomic bins was performed with MetaBAT2 (9). Taxonomy was assigned to metagenomic bins using GTDB-Tk (10). The completeness and duplication of bins were assessed using BUSCO (11) and QUAST (12). For each sample and assembly algorithm, taxonomy at the species level could be assigned to only a single bin, with all remaining bins being highly incomplete and likely not representing distinct pathogen species in the input sample. Downstream analysis was performed on whole-metagenome assemblies. For each sample, the metagenome assembly (produced by either SPAdes or MEGAHIT) with the highest BUSCO completeness and lowest BUSCO duplication values at the domain level was selected. The selected metagenome assemblies were uploaded to the AREScloud Web application (release 2022-01; Ares Genetics GmbH, Vienna, Austria) for genomic prediction of antimicrobial susceptibility. The platform used stacked-classification machine-learning (ML) WGS-AST models trained on ARESdb (13), combined with rule-based resistance prediction via ResFinder 4 (14), to provide species-specific susceptibility/resistance (S/R) predictions. AST predictions for a total of 17 antibiotic compounds were generated for samples belonging to six of the most common hospital-acquired pathogens. True negatives (TN) were defined as data points where both the reference method (phenotypic AST) and the test method (AREScloud) returned a negative (i.e., susceptible) result, true positives (TP) were defined as data points where both methods returned a positive (i.e., resistant) result, false positives (FP) were defined as data points where the reference method returned a negative result and the tested method returned a positive result, and false negatives (FN) were defined as data points where the reference method returned a positive result and the tested method returned a negative result. Very major error (VME) and major error (ME) rates were defined according to CLSI M52 guidelines (15) as the fraction of cases identified as being resistant by the reference method that were identified as being susceptible by the tested method [FN/(FN +TP)] and the fraction of cases identified as susceptible by the reference method that were identified as being resistant by the tested method [FP/(FP + TN)], respectively. The categorical agreement (CA) between the results of WGS-AST and conventional AST was calculated [CA = (TN + TP)/(TN + FP + FN + TP)] for antimicrobial-organism combinations. In silico detection of resistance genes was performed by screening the genome assemblies for each isolate against the NCBI resistance gene database (16, 17) using abricate v0.9.8 (https://github.com/tseemann/abricate) with default parameters.

Real-time PCR for resistance genes.

The utility of PCR for the targeted detection of key antimicrobial resistance (AMR) genes directly from blood culture extracts was also assessed using real-time TaqMan PCR assays. In brief, reaction mixes were prepared with 10 μL QuantiTect probe PCR master mix (Qiagen), 1.0 μM primer, 0.25 μM probe, and 2.0 μL purified gDNA diluted 1:1,000 in molecular-grade water, with a total reaction mix volume of 20 μL. Assays included the detection of ERV-3 for host gDNA depletion (18), the 16S gene in microbial DNA extracts, and the following antimicrobial resistance genes: 16S methylase genes (armA, rmtF, rmtB, and rmtC) (19), extended-spectrum β-lactamase (ESBL) genes (bla_SHV-5/12, bla_VEB, and bla_CTX-M groups 1 and 9) (20, 21), carbapenemase genes (bla_KPC, bla_NDM, bla_IMP-4, bla_VIM, and bla_OXA-48-like) (22), ampC (bla_CMY-2-like) (this study) (Tables S1 and S2), and a colistin resistance gene (mcr-1) (23, 24). Reaction mixtures were run on a Rotor-Gene Q real-time PCR thermocycler (Qiagen) under cycling conditions of 95°C for 15 min, 45 cycles at 95°C for 15 s, and 60°C for 30 s. Result analysis was conducted using Rotor-Gene 6000 series software. SPS removal assays were conducted by PCR gene amplification and gel electrophoresis to assess PCR inhibition of increasing 1:10 dilutions. GoTaq reaction mix was prepared with 6.5 μL of GoTaq (Promega), 1.0 μM primers, 3.0 μL of molecular-grade water, and 1 μL of the template. Cycling conditions (C1000 Touch thermocycler; Bio-Rad) were 95°C for 5 min; 35 cycles at 95°C for 40 s, 55°C for 1 min, and 72°C for 1 min; and 72°C for 5 min. A 3% agarose gel with 4 μL of ethidium bromide was loaded with 5 μL of the PCR mix and 2 μL of GeneRuler 100-bp ladder Plus (Thermo Fisher) and run for 40 min at 100 V. Gel images were captured using a ViberLourmat UV Gel Dock instrument.

Commercial rapid diagnostic instruments.

For comparison, positive blood culture broth samples were also tested using two commercially available platforms that provide rapid species identification and limited AMR gene profiling (blood culture identification panel [BCID] on the BioFire FilmArray instrument; bioMérieux) as well as predictive MICs using morphokinetic cellular analysis (Accelerate Pheno; Accelerate Diagnostics). The BCID and Accelerate Pheno tests were performed according to the manufacturers’ instructions, with the exception of blood culture transfer to the BCID testing pouch using a 27 1/2-gauge needle (Henke Sass Wolf) and a syringe.

Conventional AST.

All molecular, rapid diagnostic, and genome-based identification/AST tests were compared to conventional culture-based methods validated for clinical use at Pathology Queensland for diagnostic testing. Species identification was performed using matrix-assisted laser desorption ionization–time of flight (MALDI-TOF) mass spectrometry (Vitek MS; bioMérieux) on pure cultured isolates, with AST being performed by Vitek 2 automated broth microdilution (N-246 AST cards; bioMérieux), using EUCAST clinical breakpoints applicable at the time (2019). For certain species (e.g., Campylobacter jejuni), AST was undertaken using disk diffusion according to EUCAST methods (25). Conventional testing was considered the standard against which molecular and genomic tests were compared.

Ethics.

This study was approved as a low- to negligible-risk study with waiver of consent by the Royal Brisbane and Women’s Hospital and ratified by the UQ Human Research Ethics Committee (LNR/2018/QRBW/44671).

Data availability.

Raw sequence reads have been uploaded to the NCBI database under BioProject accession number PRJNA877287. Taxonomic classifications for blood culture samples can be visualized at https://fordegenomics.github.io/detect.

RESULTS

Blood culture microorganisms.

Of the 37 positive blood culture broth samples tested across all platforms, 35 had monomicrobial growth according to conventional identification methods, including Escherichia coli (n = 14), Klebsiella pneumoniae (n = 2), Klebsiella oxytoca (n = 1), Enterobacter hormaechei (n = 1), Morganella morganii (n = 1), Proteus mirabilis (n = 3), Pseudomonas aeruginosa (n = 5), Pseudomonas mosselii (n = 1), Bacteroides thetaiotaomicron (n = 1), Campylobacter jejuni (n = 2), Elizabethkingia anophelis (n = 1), Yokenella regensburgei (n = 1), Pasteurella multocida (n = 1), and Vogesella perlucida (n = 1). Two samples gave polymicrobial results by conventional culture testing (E. coli and Enterococcus faecium; E. coli and K. pneumoniae) (see Data Set S1 in the supplemental material).

DNA purity.

DNA quality and purity were assessed from host depletion to subsequent microbial DNA extraction (Table S3). The MolYsis Complete kit extraction process includes host depletion and microbial DNA extraction. Extraction concentrations were less than 1 ng/μL, with poor A₂₆₀/A₂₃₀ ratios, indicating potential extraction salt carryover. The MolYsis kit was consequently used for host depletion before microbial DNA extraction with the minikit or the UltraClean kit. The minikit method had lower yields of DNA than the UltraClean method, which was also of poor purity, with an A₂₆₀/A₂₃₀ ratio of 0.6, an adequate A₂₆₀/A₂₈₀ ratio of 0.3, and low quantification DNA ratios. The purity was improved with SPRI bead cleanup, which removed contaminants and/or inhibitors and provided the opportunity to increase the concentration through low-volume elution (Table S3). An alternative method of ethanol precipitation was trialed to concentrate DNA; however, DNA was continually lost, and the method was abandoned. Another process to remove inhibitors was through the Qiagen DNeasy PowerClean column; purified DNA was eluted, but there was a 50 to 80% loss in DNA concentrations (results not shown). The UltraClean kit has an inhibitor removal reagent that is effective at purifying the DNA and, with the amount of starting material, eluting high concentrations for downstream applications.

DNA length, size selection, and concentration.

Depending on the downstream application, the DNA length ranged from >30 kbp with minikit extraction to an average of 16 kbp with the UltraClean kit (Table S3). The application of the Circulomics short-fragment eliminator removed smaller fragments while maintaining the input DNA concentration of longer fragments. Size selection with SPRI beads at a ratio of 1:1.5 removed smaller fragments of around 4 kbp, without shearing longer fragments resulting from either extraction method. To create smaller gDNA fragments, extracts were diluted to 150, 100, and 50 μL and sheared for a fragment size of 10 kbp, with resulting sizes of 11,048 to 11,867 bp.

Downstream molecular applications.

Forty direct BC extracts, 20 from each extraction method, that underwent short-read sequencing were analyzed for taxonomy and predictive AST profiling as well as real-time PCR for AMR gene detection. For extensive investigation of applications, one blood culture sample (DETECT-110 [containing E. coli carrying bla_CTX-M-15]) was extracted by both methods, along with the cultured isolate. Another blood culture sample (DETECT-112 [containing non-ESBL-carrying E. coli]) was used for time and temperature experiments, and finally, one additional blood culture that contained the reference E. coli strain EC958 with a known number of plasmids was tested (Table 1). Short-read sequencing was possible with both extraction methods, and AMR profiling confirmed 6 genes with the same identity; this was validated by real-time PCR for the bla_CTX-M-15 gene as well as the expected absent genes. Identical AMR gene profiles were verified with long-read sequencing by both extraction methods using the transposase library preparation, also observed with the UltraClean and ligation library preparations. Mini-Pure extraction with the ligation library preparation and long-read sequencing resulted in a loss of genes. E. coli EC958 was sequenced using UltraClean DNA extraction (26). With no size selection and both the transposase and ligation library preparations, all plasmids were accounted for (Table 1).

TABLE 1.

Downstream molecular applications^e

Method	Molecular applicationc	Assay type	AMR gene detected (CT)	Taxonomy	No. of plasmids or genesd
Plasmids
UltraClean extraction with isolate EC958 (QC strain)	Short-read sequencing	Nextera Flex	Yes	E. coli	2
	Long-read sequencing	Transposase	Yes	E. coli	2
		Ligation	Yes	E. coli	2
Genes
UltraClean extraction with BC containing CTX-M ESBL-carrying E. colia	qPCR	TaqMan probe	Yes (22.0)
	Short-read sequencing	Nextera Flex	Yes	E. coli	6
	Long-read sequencing	Transposase	Yes	E. coli	6
		Ligation	Yes	E. coli	6
UltraClean extraction with BC containing CTX-M ESBL-carrying E. colia	qPCR	TaqMan probe	Yes (23.0)
	Short-read sequencing	Nextera Flex	Yes	E. coli	6
	Long-read sequencing	Transposase	Yes	E. coli	6
		Ligation	Yes	E. coli	6
Mini-Pure extraction (+ SPRI bead purification) with BC containing CTX-M ESBL-carrying E. colia	qPCR	TaqMan probe	Yes (24.2)
	Short-read sequencing	Nextera Flex	Yes	E. coli	6
	Long-read sequencing	Transposase	Yes	E. coli	6
		Ligation	Yes	E. coli	5
Duration study (1–5 days) using non-ESBL-carrying E. colib	qPCR	TaqMan probe	No CTX-M to detect
	Short-read sequencing	Nextera Flex	Yes	E. coli	6
	Long-read sequencing	Transposase	Yes	E. coli	6
Freeze-thaw using non-ESBL-carrying E. colib	qPCR	TaqMan probe	No CTX-M to detect
	Short-read sequencing	Nextera Flex	Yes	E. coli	6
	Long-read sequencing	Transposase	Yes	E. coli	6

^aSample DETECT-110 (BC sample containing ESBL-carrying E. coli [bla_CTX-M-15]).

^bSample DETECT-112 (BC sample containing non-ESBL-carrying E. coli).

^cqPCR, quantitative PCR to detect bla_CTX-M-15.

^dDETECT-110 contained the following AMR genes: aadA5, bla_CTX-M-15, bla_EC-5, dfrA17, mph(A), and sul1. DETECT-112 contained the following AMR genes: aph(3″)-Ib, aph(6)-Id, bla_EC-5, bla_TEM-1, dfrA5, and sul2 [note that the tet(34) gene was found with only 73% coverage, so it is not included in the number of AMR genes].

^eDownstream molecular applications included qPCR and short- and long-read sequencing from direct BC extracts and cultured isolates using two methods, including a quality control (QC) strain (EC958), and studies of the effects of bench time and temperature.

Taxonomic identification of metagenomic samples and in silico predictive AST.

Taxonomic identification of metagenomic samples down to the species level yielded good agreement with the results of conventional testing; for two samples, agreement to only the genus level was achieved (Vogesella urethralis by metagenomics, identified by Vitek MS as Vogesella perlucida). For one sample, metagenomic analysis identified Shigella flexneri, which was identified by Vitek MS as E. coli. MALDI-TOF methods can be unreliable in discriminating E. coli from Shigella spp. (27); however, additional phenotypic testing on this isolate (including negative Shigella-specific agglutination tests) confirmed the identification as E. coli. In two polymicrobial samples, no presence of a second species was found during the processing of metagenomic reads, with only E. coli and S. flexneri being identified in each sample by metagenomic sequencing (Data Set S1). Interestingly, in one of the polymicrobial samples, the Accelerate Pheno and BCID systems detected E. faecium but not E. coli. DNA extraction using the UltraClean kit from a blood culture negative (no-growth) control yielded minimal reads containing microbial DNA (with 97% of reads mapping to Homo sapiens). Full details of taxonomic classifications for each sample (including negative controls, freeze-thawed samples, and samples kept on the bench for up to 5 days) can be found at https://fordegenomics.github.io/detect.

The performances of the whole-genome sequencing–antimicrobial susceptibility testing (WGS-AST) models were assessed for a subset of 6 common pathogens and 17 antibiotic compounds. The overall categorical agreement (CA) was 95%, with 11% very major errors (VME) (false prediction of susceptibility) and 3.9% major errors (ME) (false prediction of resistance) (Tables 2 and 3; Table S4). The CA was >95% for 5/6 of the common bloodstream pathogens (E. coli, K. pneumoniae, P. mirabilis, P. aeruginosa, and C. jejuni), while it was lower for K. oxytoca (66.7%), reflecting errors in predicting ceftriaxone susceptibility, likely due to the challenge of chromosomal and inducible cephalosporinase genes such as bla_OXY-2 (28).

TABLE 2.

Performance of WGS-AST for the 6 most common Gram-negative pathogens^a

Taxon	Categorical agreement (%)	VME (%)	ME (%)	No. of samples
				TN	FP	FN	TP
All	95	11	4	269	11	4	31
Escherichia coli	96	0	5	187	9	0	26
Klebsiella pneumoniae	96		4	25	1	0	0
Pseudomonas aeruginosa	97	0	3	33	1	0	1
Proteus mirabilis	100	0	0	16	0	0	2
Klebsiella oxytoca	67	67	0	6	0	4	2
Campylobacter jejuni	100		0	2	0	0	0

^aVME, very major errors; ME, major errors; TN, true negative; FP, false positive; FN, false negative; TP, true positive. Blank cells indicate insufficient data.

TABLE 3.

Performance of WGS-AST for Gram-negative-active antibiotics across the species tested^a

Compound	Categorical agreement (%)	VME (%)	ME (%)	No. of samples
				TN	FP	FN	TP
All	95	11	4	269	11	4	31
Amikacin	100		0	22	0	0	0
Amoxicillin-clavulanic acid	82	20	17	10	2	1	4
Ampicillin	100	0	0	6	0	0	10
Cefazolin	71	0	33	10	5	0	2
Cefepime	100	0	0	18	0	0	1
Cefoxitin	93		7	13	1	0	0
Ceftazidime	96	0	5	19	1	0	2
Ceftriaxone	95	50	0	18	0	1	1
Ciprofloxacin	96	0	4	25	1	0	1
Gentamicin	100	0	0	24	0	0	1
Meropenem	100		0	22	0	0	0
Norfloxacin	100	0	0	13	0	0	1
Piperacillin-tazobactam	100	0	0	15	0	0	1
Sulfamethoxazole-trimethoprim	90	25	6	15	1	1	3
Ticarcillin-clavulanic acid	67	100	0	2	0	1	0
Tobramycin	100	0	0	24	0	0	1
Trimethoprim	100	0	0	13	0	0	3

^aVME, very major errors; ME, major errors; TN, true negative; FP, false positive; FN, false negative; TP, true positive. Blank cells indicate insufficient data.

For exploratory research purposes, in cases where neither WGS-AST models nor species-relevant ResFinder 4 panels existed for uncommon pathogens, nonpanel ResFinder 4 calls (based solely on the presence of known AMR markers related to the compound in question, disregarding taxonomy) were used (Table S5). The resulting set of WGS-AST calls encompassed all species found across metagenomic samples. Calls produced in this way exhibited higher rates of VME of 50.1%. This was particularly the case for taxa far removed from core ResFinder target taxa, such as Elizabethkingia and Yokenella, and for combination agents such as ticarcillin-clavulanate. Specifically, out of 39 false-susceptible exploratory predictions, 22 were for combination agents, and 13 were calls for unusual species for which no WGS-AST models or species-relevant ResFinder 4 rules exist.

Commercial rapid diagnostics.

Thirty-seven samples were assessed with new commercially available rapid diagnostic platforms. Out of 35 monomicrobial samples, the BioFire FilmArray BCID panel was able to identify 4 pathogens at the genus level and 22 at the species level, although in 2 cases, it gave an incorrect dual identification, with Proteus being identified as a second genus in 2 samples harboring E. coli and K. pneumoniae, respectively. A similar performance was shown by Accelerate Pheno, which identified 7 pathogens at the genus level and 18 at the species level. In 8 cases, the two instruments gave no identification, and these all included off-panel pathogens; in only 1 case was a discordant result observed (Yokenella regensburgei misidentified as Enterobacter spp.). Out of 2 polymicrobial samples, both instruments correctly identified only 1 of the 2 cultured pathogens (E. coli and Enterococcus spp., respectively). In one case, the Accelerate Pheno run failed on a sample harboring E. coli, while BCID failed 3 times, but when repeated, it was able to give correct results. The agreement of AST results according to Accelerate Pheno and conventional testing was 97.5% (272/279 susceptibility tests performed). Overall, most of the disagreement of Accelerate Pheno with conventional testing was observed for amoxicillin-clavulanate susceptibility (3/19 cases), with Accelerate Pheno reporting 3 E. coli isolates as resistant that tested susceptible by Vitek 2. The times to results for different diagnostic tests are shown in Fig. 2.

The average time to BC positivity for our samples was 16.1 h. The turnaround time of BCID for pathogen identification and antimicrobial resistance gene detection is 1 h from blood culture positivity, while the turnaround times of Accelerate Pheno are 1 h for pathogen identification and 7 h for AST. If implemented in a clinical laboratory, pathogen identification and WGS-AST based on short-read direct metagenomic sequencing would be available at approximately the same time as results based on conventional testing. The time from fastq file uploading to WGS-AST results via AREScloud is approximately 1 h for metagenomic samples, including multiple samples run in parallel.

DISCUSSION

A major barrier to direct sequencing from blood samples to detect pathogenic bacteria is the limited sensitivity in patients with low loads of bacterial DNA in the blood at the time of presentation (29). The addition of a culture amplification step by sequencing from positive blood culture broth leads to a significant improvement in the amount of bacterial DNA available for sequencing. Currently, no protocols are available for host-depleted genomic DNA extraction directly from blood culture broth, which is also suitable for multiple downstream molecular applications. These protocols often require the removal of host genomic DNA and/or microbial elution at high yields. The MolYsis Complete and Basic kits effectively removed host DNA. The MolYsis Complete kit included host depletion and microbial extraction; however, the DNA yield was too low for the starting DNA input for long-read sequencing, whereas the Illumina Nextera Flex library prep kit permits an input of as low as 1 ng/μL, with an adjustment to the amplification cycle step. The Basic kit uses host depletion only and is ideal for pelleting microbial cells for alternative extraction methods. The subsequent extracts from both the Mini-Pure and UltraClean methods had increased yields suitable for subsequent applications.

Previous studies utilized primarily spiked blood culture broth and extracted DNA without nontarget human “host” depletion (3, 30–33). Host depletion is important for several reasons: there may be ethical restraints for the sequencing of human DNA, it mitigates inefficiencies when over 80% of reads comprise off-target human DNA, and it may optimize sensitivity for direct microbial DNA analysis. There have been a variety of commercial total DNA extraction kits reported for specific molecular assays. This study optimized a two-kit method with minimal out-of-kit modifications, resulting in quality concentrated DNA for PCR and sequencing applications. The final DNA eluate was host depleted, with the removal of heme, the SPS preservative (which acts as a PCR inhibitor), and other agents that neutralize antibiotics, while microbial DNA was concentrated, minimally sheared, and eluted in a non-EDTA buffer (4, 6, 34).

In an effort to ensure high DNA yields and options to increase low-yield extractions as well as remove short DNA, kit protocol modifications or commercial kits were evaluated. The QIAamp Mini-Pure kit was modified to concentrate DNA initially from a two-step 2× 50-μL to a 2× 25-μL elution, with an optional SPRI bead clean-up step eluting a smaller volume. This resulted in pure DNA at suitable concentrations for PCR and library preparation. The SPRI bead clean-up and Circulomics short-read eliminator protocols favored size selection and successfully removed short fragments of DNA (~4 kbp) appropriate for long-read sequencing to prevent fuel usage in flow cells, although the removal of short DNA may lead to the loss of resistance plasmids (35). DNA shearing by the g-TUBE (Covaris) technique produced DNA fragments suitable for sequencing long-read ligation library preparations (36, 37).

Plasmid recovery by the extraction methods was assessed by the downstream sequencing of a fully annotated reference genome (E. coli EC958), which contains two plasmids (26). QIAamp Mini-Pure enzyme-based extraction will not cleave plasmids and, in turn, provides no ends for ligation library preparation (35). Although not undertaken, DNA shearing by g-TUBE or Megaruptor techniques produces DNA fragments suitable for sequencing library preparations (36, 37). The UltraPure method using mechanical and chemical lysis was effective at producing optimal DNA lengths with available ends. The results of sequencing analysis correlated with the published EC958 sequence, with the accurate number of plasmids being identified.

The two optimized extraction protocols were effective at the removal of BC preservatives and PCR inhibitors, including SPS and other known inhibitors such as antibiotic neutralizers, as verified by the PCR dilution experiments and purity data. The Qiagen DNeasy PowerClean column was investigated and resulted in pure DNA; however, there was a consistently observed 50 to 80% loss of the DNA yield, too low for long-read sequencing. The DNA preservative EDTA affects enzymatic activity during PCR and was replaced in the minikit with a non-EDTA buffer (6, 38). The success of the pure, inhibitor-free DNA was validated by PCR, and there was no observed impact on DNA library preparation with the Illumina Nextera Flex transposase tagmentation process or Nanopore transposase or ligation methods.

The quality and purity of DNA were consistent with the optimized protocols and with the duration and temperature of storage of the BC. To cover the possibility that a BC was unable to be extracted upon flagging positive, duration and storage temperature experiments were performed. The BCs were extracted on day 1 as the baseline and then every 24 h up to day 5. Furthermore, an aliquot of the BC was frozen at −80°C, and extraction was performed on day 5. In comparison to day 1, the quality and purity were maintained for each day and from the frozen extracts; short- and long-read sequencing analysis confirmed identical AMR gene profiles.

Correct species identification and detection of AMR genes from metagenomic data derived from clinical samples are critical steps in the application of direct sequencing for infectious disease diagnostics. However, the correlation between the presence/absence of AMR genes and the resistance phenotype to guide appropriate antibiotic therapy is not straightforward (39). We employed a machine-learning algorithm based on a sample bank with matched whole-genome-sequenced clinical isolates and AST results collected from several international centers (13). Our data show that phenotypic predictions from metagenomics data can be reliable for the most common Gram-negative pathogens encountered in patients presenting with bloodstream infections. However, training data sets will need to include a larger number of rarer pathogens or resistance phenotypes before reliability can be ensured in these infrequent cases. The antibiotic agent for which WGS-AST was the least reliable was ticarcillin-clavulanate, but this agent is not widely used in current practice (and is not commercially available in Australia, for instance). While the use of direct metagenomic sequencing and WGS-AST from positive blood culture broth holds promise, current methods using Illumina short-read sequencing remain time-consuming, would offer few time advantages over conventional methods, and would be slower than other emerging rapid diagnostics, including those with predictive MICs (such as Accelerate Pheno). However, it is likely that other sequencing platforms, such as Oxford Nanopore, may be able to reduce the time to sequencing results, and this needs ongoing evaluation. The application of direct sequencing from blood cultures may also hold promise for the accelerated identification of slow-growing, antibiotic-affected, or fastidious organisms or where conventional phenotypic methods take days to complete.

Limitations to this study are acknowledged. We used only a limited number of samples, and while a range of organisms were prospectively collected and included common Gram-negative species, a broader range of pathogens, including diverse AMR phenotypes as well as Gram-positive organisms, would need to be assessed to understand the reliability and clinical utility of this approach. While we assessed the utility of DNA extraction methods for a variety of molecular applications, including long-read Nanopore technology, we used only Illumina short-read sequencing on all samples. While Illumina sequencing is of high fidelity, is a very reliable WGS method, and is increasingly becoming available in clinical laboratories, it can be slower than Nanopore sequencing, which can return sequencing results in real time. Further studies to reduce the time to results with Nanopore sequencing from blood culture samples are warranted.

In summary, the UltraClean method proved optimal for host-depleted, microbially enriched, inhibitor-free DNA extraction and downstream molecular applications. Through one extraction, there is the ability to use DNA for PCR and short-read and long-read sequencing without plasmid loss. These methods support the development of molecular diagnostic assays and metagenomic sequencing directly from blood cultures, leveraging the advantages of preenrichment through culture amplification using commercial blood culture systems that are widespread in clinical practice. We have also demonstrated the utility of the use of machine-learning algorithms directly on clinical samples to accurately define effective antibiotic therapy. Further validation work and, ultimately, evaluations of the clinical benefit of such approaches are warranted.