Four strains representing both Gram-positive and Gram-negative bacterial species were selected for this study: Campylobacter coli GPT20-006, Escherichia coli GPT22-004, Enterococcus faecalis GPT22-006, and Staphylococcus hominis NT-2025, as described in our previous study (19). These strains had previously been sequenced with PacBio or hybrid assembly with ONT and Illumina (hybrid assembly), resulting in complete circular reference genomes as described in Table 1. Genomic DNA (gDNA) was extracted using the DNeasy Blood and Tissue spin-columns kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol with modifications: all post-lysis vortexing steps were replaced with flicking the tube and pipetting gently up-and-down to mix, and 15 mg/mL lysostaphin was added for lysis of the S. hominis strain.

To evaluate the effect of multiplexing level and DNA input concentration, nine run configurations were generated by combining three DNA concentrations (5 ng/μl, 10 ng/μl, and 20 ng/μl) with three multiplex levels (12-, 24-, and 36-plex). For each run configuration, 10 μL of gDNA was used per sample, corresponding to final input amounts of 50 ng, 100 ng, and 200 ng measured with Qubit 4 Fluorometer (Invitrogen, Thermo Fisher Scientific, Waltham, MA, USA) using the Qubit High Sensitivity Quantification Kit (Invitrogen, Thermo Fisher Scientific), except for concentrations >120 ng/μl where the Invitrogen Qubit Broad Range Quantification kit (Invitrogen, Thermo Fisher Scientific) was used according to the manufacturer’s protocol.

Each strain was represented across all three DNA concentrations and included in multiple technical replicates within each multiplexing level. Library preparation was performed using the Rapid Barcoding kit v14 (SQK-RBK114.96).

The prepared libraries were loaded onto R10.4.1 flow cells (FLO-MIN114) with more than 800 active pores pre-sequencing. Flow cells were sequenced using ONT’s MinION or GridION sequencing platforms with the MinKNOW v23.11.4 software. Basecalling was performed using the super-accurate basecalling (SUP) model, and reads were filtered based on a minimum Phred quality score (Q-score) > 10 and read length > 400 bp thresholds.

Genomes were assembled with Flye v2.9.1 (20) using default settings, and per-sample read metrics were generated with NanoStat v1.6.0 (21). As in our previous study (19), MLST profiles and AMR genes (ARG) were identified for the reference genomes and assemblies using our previously described in-house pipeline (22), which employs MLST v2.0 (22) with 100% identity and 100% coverage, and ResFinder v4.2.3 (database version 2.3.2) (23, 24) with ≥ 95% identity and ≥ 95% coverage.

R statistical software v4.4.2 (25) in RStudio (26) was used for all statistical analyses and all plots were visualized using the ggplot2 package (27). To assess the influence of the run configuration on the WGS performance, Firth-penalized logistic regression was conducted using the logistf package (28). The model evaluated the effects of multiplexing level and input DNA concentration on four binary outcomes: (a) achieving ≥ 30 × sequencing depth, (b) complete MLST locus detection, (c) full ARG detection, and (d) overall WGS success, defined as meeting all three criteria. The model incorporated an interaction term between multiplexing level and DNA input concentration and was adjusted for bacterial strain to account for strain-specific variability. The model fit was assessed using likelihood ratio test, and odds ratios (OR) with 95% confidence intervals (CI) reported relative to the reference condition (12-plex × 100 ng input DNA). Statistical significance was defined as p < 0.05. For each run configuration, the binary outcomes were summarized as n/N (%), where n is the number of samples meeting the criterion, and N is the number evaluated. The workflow from sequence data generation to analysis and how WGS success was defined can be found in Supplementary Figure S1.

All cost estimates were reported in Euros (€), exclusive of value added tax (VAT). Costs associated with auxiliary equipment and personnel were excluded from the cost analysis. The cost per sequencing run was calculated based on prices from Danish suppliers as of January 2025 and reported both by including and excluding DNA extraction, covering the process from library preparation to finalized WGS data. Per-sample costs were calculated for each multiplexing level. Costs for ONT flow cells were derived from single-unit purchase prices. The library preparation was estimated using the Rapid Barcoding kit 96 v14 and included additional consumables not supplied with the kit (i.e., pipette tips, tubes, and chemicals). Cost per sample was determined for each specific multiplexing level.

The ONT runs yielded between 2.2 and 14.61 gigabases (Gb) of data with a median output of 9.65 Gb. The highest data outputs were observed in the three runs using 100 ng input DNA with the 12-plex, 24-plex, and 36-plex configurations, which generated 13.85 Gb, 14.61 Gb, and 13.06 Gb, respectively, with 64.80 to 72.05% of reads passing quality thresholds of Q-score ≥ 10 and read length ≥ 400 bp, respectively. In comparison, the 12-plex and 24-plex runs using 50 ng input DNA produced 9.65 Gb and 11.29 Gb, respectively. Conversely, the 24-plex and 36-plex runs with 200 ng input DNA yielded less data, 3.38 Gb and 2.20 Gb, respectively, although 71.93 to 75.56% of reads passed quality filtering (Figure 1).

The cost per ONT sequencing run, including DNA extraction, ranged from €791.35 to €803.23. Cost per sample varied substantially depending on multiplexing level: 12-plex runs were the most expensive at €65.95 per sample (or €72.05 when including DNA extraction), followed by 24-plex at €33.22 (€39.32 with extraction), and 36-plex €22.31 (€28.41 with extraction). Thus, increasing multiplexing from 12 to 36 samples reduced the per-sample cost by approximately 60%, nearly a threefold difference (Table 2). Across all run configurations, flow cell expenses accounted for the largest proportion of the cost, ranging from 68.44 to 80.97% of the total per-sample cost.

Our study indicates the interaction between multiplex level and input DNA influenced WGS success significantly. The results show a clear interaction between multiplexing level and DNA input. Runs with 12-plex and 24-plex configurations using ≤100 ng DNA consistently achieved high sequencing depth, accurate MLST assignment, and robust ARG detection. In contrast, 36-plex runs showed reduced success, particularly when combined with 200 ng input DNA, where overall WGS success dropped to 2.8%. The 24-plex and 36-plex runs with 200 ng input DNA yielded a lower data output (≤3.4 Gb data) than the other run configurations in the study (≥6.9 Gb data) (Figure 1). Interestingly, reducing DNA input to 50 ng did not negatively affect performance; in fact, 12-plex × 50 ng achieved 100% success, and 24-plex × 50 ng maintained >80% success (Figures 2a–d).

These findings suggest that lower DNA loading can mitigate some negative effects of high multiplexing, whereas excessive DNA input combined with large batch sizes significantly compromises sequencing efficiency. This supports previous protocols recommending 100 ng input DNA for multiplexing >4 samples (29), however, does not align with ONT’s own protocols.

Our findings align partially with other studies and protocols. Landman et al. (17) recommends multiplexing maximum 16–18 bacterial samples for AMR surveillance using the ONT SQK-RBK-114 library preparation kit and R10.4.1 flow cells, whereas other studies have successfully multiplexed 10–12 bacterial samples for species such as Clostridioides difficile and Salmonella (13, 30). ONT’s own NO-MISS protocol (ISO_9205_v114_revG_09May2025) describes a flexible range of multiplexing 4–24 bacterial isolates per flow cell. The reduced WGS success observed in our 36-plex runs supports these guidelines, indicating that multiplexing beyond 24 samples may compromise data quality, particularly when paired with high DNA input.

The 24-plex and 36-plex runs with 200 ng input DNA yielded a lower data output (≤3.4 Gb data) than the other run configurations in the study (≥6.9 Gb data) and a reduced WGS performance was observed with high multiplexing level (36-plex) and input DNA (200 ng) (Figures 1, 2).

It is important to note that the DNA extraction method and library preparation kit used in this study were deliberately selected to maintain feasibility in low-resource laboratory settings. The DNeasy Blood & Tissue spin-columns kit was selected because it is widely available in LMICs, and is cost-efficient, time-efficient, and generally robust, despite not being specifically optimized for high-molecular weight (HMW) DNA. The protocol does not include any RNA removal steps or size selection. Incorporating such steps would require additional reagents or separate kits, thereby increasing both cost and logistical burden and reducing overall cost-effectiveness (19). Similarly, the SQK-RBK114 kit employs a transposase-based fragmentation and barcoding step without purification to remove short DNA fragments, instead prioritizing simplicity and speed over long read lengths and yield. In contrast, ONT’s ligation-based library preparation kits incorporate multiple purification steps that remove short fragments, thereby enriching for HMW DNA and generally yielding higher total data output (9, 15, 31). However, small fragments will be preferentially sequenced on the ONT platform due to a higher translocation speed than HMW (32). The Nanopores wear down faster with small molecular wight DNA, thereby limiting the total amount of generated Gb yield (33). These effects could accumulate with increased input mass (both from level of multiplexing and input DNA quantity), thereby causing the reduced data yield observed for 24-plex and 36-plex runs with 200 ng input DNA. An reduced performance across run configurations may also reflect inhibitory effect from EDTA in the DNeasy Blood & Tissue elution buffer (34). Although ONT reports significant performance reduction of rapid sequencing kit at 5 mM of EDTA, even the 0.5 mM concentration in the buffer may have contributed to reduced efficiency when compounded with other factors at high multiplexing (35). A different elution buffer without EDTA could have avoided this potentially inhibitory effect, however that would have required procurement of additional reagents not included in the commercial kit. Any additional reagents were avoided, if possible, in the study to maintain feasibility in LMICs.

The reduced performance for high multiplexing and input DNA can in large part likely be attributed to the lack of additional purification steps in the DNA extraction and library preparation.

The four strains included in our study had varied genome sizes (1.8–5.1 Mb) (Table 1). If applying this workflow to WGS runs containing only large genomes (e.g., Enterobacterales), a lower sequencing depth would be generated per Gb. This is a consideration when designing a workflow usable for all World Health Organization (WHO) Global Antimicrobial Resistance and Use Surveillance System (GLASS) and the Food and Agriculture Organization of the United Nations (FAO) priority pathogens and could indicate that multiplex levels above 24 samples may be unsuitable.

Despite libraries being normalized and N50 being comparable across the four strains, the Campylobacter strain yielded fewer WGS successes, primarily due to low sequencing depth (Supplementary Table S3). This underrepresentation is unlikely to be attributed to the species’ low GC content, as the C. coli, E. faecalis, and S. hominis strains all exhibited similar levels yet achieved high WGS success. DNA fragmentation and loading bias have nevertheless been reported as challenges for Campylobacter due to its low GC content (36). Furthermore, genome size is an unlikely cause as the strain’s smaller genome size (~1.8 Mb) should theoretically correspond to a higher number of genome copies per ng/μl (Table 1). Instead, it might reflect strain-specific differences in transposase barcoding efficiency or genome copy-number effects, potentially compounded by barcode imbalance typical of high multiplexed ONT runs as seen in other studies (17, 37). ONT’s SQK-RBK kit uses the transposase-mediated fragmentation and adapter tagging (31). The lower yield for Campylobacter may reflect reduced transposase tagging efficiency, consistent with known reduction in sequencing performance associated with DNA composition, secondary structure, and integrity in transposase-based barcoding chemistries (2, 38, 39).

While the study included Gram-positive and Gram-negative strains with varied genome sizes, broader GLASS panels will include more extreme GC contents, where composition and kit-associated biases may exacerbate sequencing depth inequities in high-multiplex runs, highlighting the need for conservative depth targets (40). Many GLASS priority pathogens, particularly Enterobacterales, frequently carry ARG on plasmids and other mobile genetic elements that require sufficient depth for contextualization (41, 42). Our findings support low (12-plex) to moderate (24-plex) multiplexing strategies for GLASS implementation. Further, these findings highlight that laboratory workflow optimization is critical to ensure that genomic surveillance data remains interpretable and comparable across organisms and sequencing sites, particularly in decentralized surveillance settings, aligning with published implementation models and expert recommendations (3, 43).

Cost, infrastructure, and supply chain challenges remain the most significant barriers to WGS implementation in African laboratories (5, 44, 45). Our analysis focused exclusively on consumables (flow cells, library preparation kits, and DNA extraction) and excluded personnel, equipment, and infrastructure, as existing costing frameworks emphasise that consumables and reagents, rather than equipment purchase, are dominant recurring cost drivers for sustainable WGS implementation in LMICs (46, 47). Within this scope, ONT WGS using the DNeasy Blood & Tissue kit and SQK-RBK-114.96 library prep kit costs €865 to €1,023 per sequence run, with per-sample cost strongly influenced by multiplexing level: €72.05 for 12-plex, €39.32 for 24-plex, and €28.41 for 36-plex (Table 2). Flow cells accounted for approximately 68 to 81% of the total cost per isolate. African cost-analysis similarly reports even higher proportions (12-plex: ~86–91%, 24-plex: ~75–82%, and 36-plex: 67–76%) for similar workflows, likely due to transport and warranty cost for the flow cells (48).

However, run configurations that reduce per-sample cost at the expense of sequencing success may increase overall programme cost through repeat sequencing and delayed reporting, causing loss of actionable surveillance data. These risks are particular significant in settings such as Sub-Saharan Africa, where supply chains are characterized by long lead times and cold-chain constraints (49).

While increasing multiplex level reduces the per-sample costs, Elton et al. (50) noted that the time to results may be increased at laboratories that do not process a high number of samples, as these would need a longer period to collect enough samples to sequence. A possible mitigation strategy may be to wash and reuse the flow cells to increase cost-effectiveness whilst not compromising turnaround time or buying flow cells in bulk, which can lower costs up to ~25% (from €761.68 to €571.26 in Sudan, as of November 2025), though warranty periods, expirations dates, and procurement challenges remain significant (5, 45, 51, 52). These findings highlight that decentralized sequencing strategies must be planned with realistic expectations regarding sample throughput, turnaround time, and reagent availability, as excessive multiplexing may reduce data quality without delivering sustainable cost savings.

Overall, our findings indicate that 12-plex runs are unlikely to be cost-effective for routine implementation in low-resource settings, where affordability is particularly critical (48). Multiplexing 24 samples using the SQK-RBK114 kit with low to moderate DNA input (50–100 ng) offers the most practical balance between sequencing success and cost for decentralized laboratories. These results highlight the importance of optimizing both multiplexing level and DNA input to ensure reliable ONT WGS outcomes. The findings support decentralized AMR surveillance strategies and align with global initiatives such as GLASS. Future studies should validate these recommendations across diverse bacterial species and explore additional cost-saving measures, including washing of flow cell for reuse and alternative library preparation kits, to enhance scalability and sustainability.

Although sequencing success declined from 91.7% at 12-plex × 100 ng to 66.7% at 36-plex × 100 ng, this difference was not statistically significant (OR = 0.15, 95% CI 0.01–1.01, p = 0.052). The lack of significance likely reflects limited sample sizes rather than absence of a biological effect. A stronger and statistically significant decline was observed when both multiplexing and input DNA were increased (36-plex × 200 ng; OR = 0.02, 95% CI 0.0008–0.58, p = 0.024), reinforcing that overloading flow cells substantially reduces sequencing performance.

The study used Rapid Barcoding Kit v14 and R10.4.1 flow cells. Use of older chemistries or alternative library preparation kits may yield different optimal configurations. Additionally, inclusion of only four bacterial strains may limit its applicability to broader AMR surveillance efforts, as it does not encompass the full spectrum of species typically encountered.

To further evaluate the applications of these findings, SeqAfrica has initiated multicenter implementation studies, including a pilot comparison of decentralized ONT WGS, using the workflow optimized here, with centralized Illumina sequencing for Salmonella as a model pathogen in South Africa, alongside One Health sentinel site activities in Ghana evaluating a broader panel of local priority pathogens in the GLASS bracket. Assessment using primary isolates from clinical and One Health surveillance contexts will be essential to evaluate robustness under routine surveillance conditions. Positioned as a foundational optimization study, this work provides a framework to inform subsequent implementation studies and capacity-building efforts for decentralized AMR genomic surveillance.