Performance of molecular methods for the detection of Salmonella in human stool specimens

Background: The relationship between asymptomatic Salmonella exposure within the gastrointestinal tract and Salmonella bacteraemia is poorly understood, in part due to the low sensitivity of stool culture and the lack of validated molecular diagnostic tests for the detection of Salmonella in the stool. The study aimed to determine a reliable molecular diagnostic test for Salmonella in stool specimens. Methods: We optimised an in-house monoplex real-time polymerase chain reaction (PCR) for the detection of Salmonella ttr and InvA genes in stool by including a selenite broth pre-culture step for Salmonella before DNA extraction and validated their specificity against other local common pathogens. Then we assessed their performance against a well-validated multiplex PCR targeting the same ttr and InvA genes and against stool culture using clinical stool specimens collected from a cohort of 50 asymptomatic healthy Malawian children that were sampled at 1-month intervals over 12 months. We employed a latent Markov model to estimate the specificities and sensitivities of PCR methods. Results: Ttr and InvA primers were both able to detect all the different Salmonella serovars tested and had superior limits of detection when DNA was extracted after selenite pre-culture. T tr sensitivity and specificity for monoplex-PCR were (99.53%, 95.46%) and for multiplex-PCR (90.30%, 99.30%) respectively. InvA specificity and specificity for using monoplex-PCR was (95.06%, 90.31%) and multiplex-PCRs (89.41%, 98.00%) respectively. Sensitivity and specificity for standard stool culture were 62.88% and 99.99%, respectively. Culture showed the highest PPV (99.73%), and monoplex- ttr had the highest NPV (99.67%). Conclusion: Test methods demonstrated high concordance, although stool culture and monoplexed ttr primers had superior specificity and sensitivity, respectively. The use of selenite pre-enrichment step increased Salmonella detection rate. Taken together, molecular detection methods used here could be used to reveal the true extent of both asymptomatic and symptomatic Salmonella exposure events.


Introduction
Salmonellae cause a huge global burden of morbidity and mortality. They are globally estimated to be responsible for 300,000 deaths [1][2][3][4] . Salmonella enterica serovars Typhi and Paratyphi A are the predominant cause of invasive Salmonella infections in south and southeast Asia and cause between 129,000 to 223,000 global deaths per year 1,3,5 . In contrast, non-Typhoidal Salmonella (NTS) serovars, principally S. Typhimurium and S. Enteritidis, are a common cause of invasive disease in sub-Saharan Africa (SSA) 4,6 . In 2017, NTS caused an estimated 535,000 cases, with SSA having the highest incidence 7 . Risk factors for invasive NTS (iNTS) disease include young age, recent malaria, and advanced HIV disease. Case fatality rates for iNTS in young children, people infected with HIV, and living in the SSA region were estimated at 13.5%, 41.8%, and 15.8%, respectively 4 . This is in marked contrast to the presentation of Salmonella disease in high-income countries, where NTS typically cause a self-limiting diarrhoeal disease in healthy individuals, while bloodstream or focal infections are rare and mainly occur in individuals with specific risk factors such as diabetes, neoplastic and autoimmune disease, or immunosuppressive therapy 8 . However, it is notable in both settings that invasive NTS disease in adults and children is not always associated with diarrhoea 9 .
We previously described in under-five-year-old children the sequential development of cellular and humoral immunity against the Salmonella serovars causing iNTS disease, and that acquisition of this immunity is associated with decreasing incidence of disease 10,11 , suggesting that this immunity is protective. Previous studies have reported that healthy young children experience transient asymptomatic episodes of gastrointestinal infection with non-typhoidal Salmonella 12,13 , and we, therefore, hypothesise that episodes of asymptomatic Salmonella exposure in the healthy gastrointestinal tract during early childhood may facilitate the development of protective immunity. Balanced against this beneficial effect of exposure, diarrhoeal disease results from enteric Salmonella exposure, and invasive NTS disease also follows episodes of asymptomatic gastrointestinal exposure in susceptible children, including those with malaria or malnutrition, or immunocompromised individuals.
Elucidating the relationship between Salmonella exposure events within the gastrointestinal tract and resultant Salmonella immunity or Salmonella disease is critical for understanding iNTS disease pathogenesis. Lack of affordable and rapid diagnostic tools for the detection of bloodstream and intestinal Salmonella disease hampers our understanding of Salmonella disease epidemiology and pathogenesis. Blood culture is considered the gold standard diagnostic test for Salmonella bacteremia and is highly specific but has a number of drawbacks; poor turnaround time of between 2 to 7 days, and low sensitivity of about 20% -30% for samples collected 7 days post-infection [14][15][16] . Molecular detection of Salmonella in blood also has limited apparent sensitivity, and various assays are in development 13,17 .
Stool culture is similarly considered the gold standard test for the detection of Salmonella in the intestinal tract. However, stool culture, even for diarrhoeal disease when the bacterial load is likely to be high, has poor sensitivity (<50%), and is labour and time-consuming 18 . Real-time PCR has a short turnaround time and is potentially highly sensitive compared to standard culture, and has the capacity for automation and testing for multiple targets 19 . However, stool PCR test performance is hindered by PCR inhibitors and a large number of genetically closely related enteric bacteria. These pose a challenge in generating highly specific and sensitive primers for real-time PCR (qPCR) for Salmonella. Furthermore, a lower infective load of Salmonella colonisation during asymptomatic infection may further limit detection by PCR.
With this background, we validated an in-house monoplex qPCR method for the detection of Salmonella in stool specimens and compared them with a validated multiplex-based qPCR and standard stool culture. Both qPCR assays used primers and probes based on the Salmonella tetrathionate respiration gene (ttr) and the Salmonella invasion gene A (InvA). Stool specimens were collected from healthy, mainly asymptomatic healthy Malawian children aged 6-18 months. Assessing a diagnostic test's performance is challenging when the existing "gold standard" test being used has known low sensitivity or specificity. Statistical methods, such as the Latent Markov model, are used to assess diagnostic tests' performance without assigning a gold standard test. Since the current reference standard is known to lack sensitivity, we employed a latent Markov model to estimate the specificities and sensitivities of PCR methods without assigning a gold standard.

Description of study participants and specimens
Stool specimens collected from a longitudinal cohort of children aged 6 -18 months recruited from Zingwangwa Health Centre (ZHC) in Blantyre, Malawi, were used to compare the performance of molecular and standard culture for detection of Salmonella in stool. The primary study started recruitment in August 2013, and follow-up was concluded in December 2014 17 . Group sensitisation of the study by well-trained study nurses was done to parents or guardians of six-month-old children attending a vaccination clinic at ZHC. Individual sensitisation was also done to parents or guardian that were interested in joining the study. Children who met the inclusion criteria of being healthy were recruited into the study after obtaining consent. Children born preterm (less than 38 weeks gestation), HIV positive or HIV exposed, and those with fever >38°C or any acute illness were excluded from the study 20 .
Stool samples were collected monthly until they were aged 18 months. Stool specimens were collected in sterile and clean containers and transported to the laboratory on the same day. From 60 children recruited at 6 months of age, 10 children withdrew from the study, and 600 stool specimens were collected and tested by culture on the day of sample collection at the College of Medicine and Malawi Liverpool Wellcome Laboratory. Molecular tests were done on frozen samples that were available at the time the tests were done.
Salmonella stool culture A matchstick head-size sample of stool was inoculated in 10 ml of selenite F broth (Oxoid, UK, catalog number: 2300631) and aerobically incubated overnight at 37 °C for 18-24 hours. The top layer (1 ml) of an overnight culture was spun at 20,000 g for 5 minutes. This is a method that has been developed in our laboratory. A 1 ul loop was used to subculture Salmonella from the pellet by spreading on Xylose Lysine Deoxycholate (XLD) agar (Oxoid, UK, catalog number: 2547703) to achieve single colonies. Careful plate spreading prevents overcrowding of colonies. A single colony was picked and cultured on sheep blood agar. An aliquot of the selenite broth was also frozen for molecular detection. A single colony of presumptive Salmonella was cultured onto sheep blood agar (Oxoid, UK, catalog number: 2910831) and MacConkey agar plates (Oxoid, UK, catalog number: 2529552) and incubated aerobically at 37°C for 18-24 hours. Salmonella colonies were then distinguished from other enteric bacteria (i.e., Citrobacter and Serratia) using triple sugar iron agar (Oxoid, UK, catalog number: 1882283) and Urea agar (Oxoid, UK, catalog number: 1779617) biochemical tests. Further Salmonella identification was determined using API® 10S (bioMérieux, France, catalog number: 1007181060) according to the manufacturer's instructions.
Monoplex-qPCR ttr and InvA assay Validation of the monoplex-qPCR ttr and InvA assay. For the monoplex-qPCR, the ttr primers and probe were designed and validated by Federal Institute for Risk Assessment, Berlin, Germany, according to the published DNA sequence of the S. enterica serotype Typhimurium ttr locus for Salmonella detection (GenBank accession no. AF282268). The use of ttr was based on 21. In this study, the ttr gene's specificity was assessed using 110 Salmonella strains representing 31 serotypes, and it demonstrated 100% specificity. The InvA gene has been widely studied and used as a pan Salmonella marker. In 22, the specificity of the InvA gene was assessed using 242 Salmonella strains representing 43 serotypes. It also demonstrated 100% specificity. InvA has shown 100% specificity in several other studies. Both primers required optimisation for use in stool specimens. The DNA sequences of all the primers and probes used in this study are listed in Table 1.
Specificity of ttr and InvA primer/probe set for Salmonella compared to other local pathogens To determine the prevalence of ttr and InvA in the genomes of Salmonella serotypes, NCBI nucleotide blast was conducted (10 February 2021). Ttr and InvA nucleotide sequences were used as the query sequence against Salmonella enterica subspecies enterica (taxid:59201) genomes with a maximum target sequence set at 1000. Both primer sequences demonstrated 100% identity for most commonly isolated Salmonella serotypes (supplementary material). To determine the specificity of the primers in vitro, 9 different locally isolated and whole genome sequenced Salmonella strains and 26 pure isolates of non-Salmonella bacterial strains locally isolated from blood culture were tested using ttr and InvA primer/probe sets ( Table 2). These 9 Salmonella strains represented all known Salmonella serotypes in the MLW isolate archive. The non-Salmonella strains were chosen because they are genetically closely related to Salmonella or because their growing conditions are similar to Salmonella. These strains were collected from MLW bacterial blood culture repository. Overnight cultures of the frozen samples were made on SBA or LB agar. One colony was then cultured in liquid media. After reaching stationary growth phase, a known and matched concentration of about 10 6 CFU was used for DNA extraction using QIAamp Fast DNA Stool Mini Kit (QIAGEN, Netherlands, catalog number: 51604) but without the bead beating step. Miles and Misra technique was used for bacteria quantification.

Limits of detection in different conditions
A well-characterised invasive S. Typhimurium ST313 strain (D23580), isolated from an HIV-negative child in Malawi, and representative of our commonest invasive bloodstream infections, was used as a reference strain for determining limits of detection in varying kinds of sample 23,24 . Three types of Salmonella samples were prepared for comparison using RT-PCR; 1) pure Salmonella isolates picked from a blood agar plate, 2) Salmonella cultured in selenite broth, and 3) Salmonella spiked into stool. Salmonella stool spiking was done to determine the inhibitory effect that stool may have on the assay, affecting the limit of detection. For this, a stool sample was collected from a healthy individual and confirmed as Salmonella negative by culture. The stool sample was thereafter diluted with PBS (50% w/v) and then spiked with S. Typhimurium, D23580, at varying doses of viable bacteria. The viable dose of Salmonella was adjusted across a range from 10 0 -10 6 CFU/ml and quantified using Miles and Misra technique. DNA was extracted for RT-PCR, as above. All experiments were repeated three times on different days by the same operator. A previously-optimised in-house PCR protocol was used 17 . Briefly, the master mix for RT-PCR was prepared using predefined quantities. A total of 20μl master-mix for each sample was comprised of the following: 12.5μl Platinum ® Quantitative PCR Super Mix-UDG (Life Technologies, USA, Catalog number: 11730025), 0.10μl specific forward primer, 0.10 specific reverse primer, 0.10 specific probe (all primers and probes at 200nM), 0.05μl ROX reference dye (Life Technologies, USA, Catalog number: 12223012) at 50nM final concentration, and 7.15μl nuclease-free water. This mixture was transferred to 96-well plate PCR wells. 5μl of test DNA, positive controls DNA (DNA from D23580), technical extraction negative control, and assay negative control (UV treated water) were added in triplicates to appropriate wells containing 20ul of mastermix. The qPCR was run for 40 cycles using Applied Biosys-tems® 7500 Real-Time PCR Systems (Life Technologies, USA). The following cycling conditions were used; initial denaturation at 95°C for 1 minute, denaturation at 95°C for 15 seconds, annealing/extension at 60°C for 30 seconds, final extension: 12°C. The threshold was set in the lag phase. An assay was considered to have passed when the positive controls were positive, and both the technical extraction negative and assay negative controls were negative. Test sample cycle threshold (Ct) values were evaluated after subtracting the baseline value. Samples with cycle threshold (Ct) values of less than or equal to 35 were considered positive.

Detection of Salmonella in clinical samples using monoplex-qPCR ttr and
Detection of Salmonella using multiplex qPCR assay As a comparator, we used a well-validated TAC assay on DNA extracted from stool samples, according to the manufacturer's protocol. The customised Taqman Array Card assay developed and validated at the University of Virginia was used. The performance of the TAC method has been previously described and has now been widely used 12,25-27 . It is used to detect multiple enteric pathogens, including bacteria, viruses, protozoa, and helminths. Targets included on the TAC card for pan Salmonella detection are InvA and ttr 28 . Phocine Herpesvirus (PHhv) and MS2 targets are included as internal positive controls.
To extract total nucleic acid (TNA) from the clinical samples for TAC assay, we used QIAamp Fast DNA Stool Mini Kit (QIAGEN, Netherlands, catalog number: 51604) -the same DNA extraction kit and protocol that were used to extract wholestool DNA for the monoplex qPCR assay, with the addition of internal extraction positive controls. For TNA extraction, each sample was extracted together with internal positive controls, Phocine Herpesvirus (PHhv), and MS2. PHhv and MS2 were added to the inhibitX buffer before being added to each sample, as previously described 28 . An assay was considered to have passed when both MS2 and PhHv internal positive (amplification crossing the threshold) and negative controls (no amplification crossing the threshold line) passed and when the sample had a sigmoid curve that crossed the threshold line. A sample was Table 1. List of primers and probes sequences used in this study. Primers and probes sequences used in this study include in-house designed InvA, ttr previously validated for Salmonella detection in food, and TAC-InvA and TAC-ttr used on a wellvalidated TAC assay as pan Salmonella primers.

Primer name Primer direction Primer code/Probe description
classified as pathogen positive at a Ct value of <35. Only results for Salmonella are reported here.

Statistical analysis
Data were recorded and analysed in MS Excel (version 16.14.1 (18061302)). Sensitivities and specificities of the different PCR methods were estimated using a latent Markov model (LMM) 29 . We have previously described the LMM and various extensions that we considered for modeling longitudinal diagnostic test data 30 . We implemented the LMM within a Bayesian framework using R (version v3.5.1) and JAGS (version 4.3.0) via the rjags (version 4.6) R package 31 . LMMs have been extensively used for discrete-time longitudinal data in the absence of a gold standard diagnostic procedure 32,33 . LMMs consists of a process model for a latent condition (in our case, the unobserved true infection status) evolving over time and a measurement model for the observed outputs (in our case, the results from the 5 diagnostic test methods) conditional on the latent state. We considered several LMMs, with and without mixed effects and with either time-homogeneous or time-heterogeneous transition matrices 30 . Convergence and identifiability of the LMM were checked by inspecting trace plots and computing Gelman-Rubin potential scale reduction factors 34 . The more complex models exhibited poor mixing or convergence of MCMC chains (most likely due to the sparse number of positive samples). As a result, the LMM we used for this dataset is a basic LMM with no random effects and a timehomogeneous transition matrix. To report positive predictive values (PPV) and negative predictive values (NPV), we calculated an estimate of the infection prevalence. For the Bayesian LMM, we report maximum a posteriori (MAP) parameter estimates together with 95% credible intervals (Crl),  specifically the highest posterior density intervals (HDI) with 95% coverage. All other analyses report (frequentist) parameter estimates and corresponding 95% confidence estimates (CI).

Ethical considerations
Ethical approval for this work was granted by the University of Malawi, College of Medicine Research Ethics Committee (P.01/13/1327). Written informed consent was obtained from the parent or guardian of each participating child.

Results
Ttr and InvA primers for Salmonella do not cross-react with closely related enteric micro-organisms We first validated the ttr and InvA primers that were used in the monoplex-qPCR assay by assessing the sensitivity and specificity of the primers for Salmonella, using a standardised number of 10 0 -10 6 CFU/ml of 9 different locally-relevant Salmonella strains and 26 non-Salmonella bacterial strains as indicated in Table 2. We included 17 strains of E. coli because of the close genomic relatedness of Salmonella and E. coli. Bacterial isolates enriched in Selenite F broth (referred here as selenite sub-cultured) or not (referred here as direct culture) were used in this evaluation. We found that ttr and InvA assays both achieved 100% sensitivity and specificity either as direct isolates or selenite sub-cultured isolates. Table 2 demonstrates that all Salmonella strains tested positive with both monoplexed primer pairs, and all other bacterial strains were negative, confirming a lack of cross-reactivity. Additionally, the NCBI nucleotide blast demonstrated that the genomic prevalence of ttr and InvA genes is high. Both primer sequences demonstrated 100% identity for most commonly isolated Salmonella serotypes.
Selenite broth culture enhances the detection of Salmonella in stool using either ttr or InvA primers The limits of detection (LOD) of qPCR for Salmonella were then determined using S. Typhimurium strain D23580 serially diluted and tested as direct isolates, selenite broth cultured samples, or isolates spiked into a culture-negative stool specimen. We found that limits of detection for ttr were 1, 10, and 100 CFU/ml, and for InvA were 1, 100, and 100 CFU/ml for selenite sub-cultured broth, direct isolates, and stool-spiked isolates, respectively, with 98.5% qPCR efficiency for ttr and 97.2% qPCR efficiency for InvA. No statistically significant difference was observed in the LOD when ttr was compared with InvA in either direct isolates (p = 0.3212), selenite sub-cultured samples (P = 0.2534), or salmonella spiked stool samples (P = 0.2361). Importantly, we found that the ttr assay was significantly different when direct isolates (LOD = 10 CFU/ml) were compared with selenite sub-cultured samples (LOD = 1 CFU/ml) (p<0.0001), and when selenite sub-cultured isolates were compared to Salmonella spiked stool (p <0.0001), and there was no significant difference when direct isolates were compared to Salmonella spiked stool (p=0.2965).
Similarly, we found that detection in InvA qPCR assay direct isolates was significantly different compared to selenite broth cultures isolates (p < 0.0001), and selenite subculture isolates were also significantly different from Salmonella spiked stool (p < 0.0001). In contrast, there was no significant difference between direct isolates compared to Salmonella spiked stool samples (p = 0.2862). In summary, we found that selenite broth overnight liquid culture of stool samples enhanced the molecular detection of Salmonella using either ttr or InvA primers, even if culture of the broth remained negative.
Ttr and InvA primers had both high specificity and sensitivity rates, while stool culture had high specificity but low sensitivity The samples from healthy children were used to determine the performance of stool culture, monoplex ttr, monoplex InvA, multiplex TAC ttr, and multiplex TAC InvA. Standard stool culture was performed on a total of 600 specimens at different time points. Molecular tests were used to detect Salmonella in the available 421 stool DNA specimens. We detected Salmonella in 23, 40, 29, 56, and 47 of 421 stool specimens, using standard stool culture, ttr, InvA, TAC-ttr, and TAC-InvA respectively. Of the 23 Salmonella stool culture-positive samples, 21 samples were also positive with either one or more molecular tests, while 2 were molecular tests negative.
High negative and positive concordance for stool culture, monoplex ttr, monoplex InvA, Multiplex ttr, and multiplex InvA Next, we explored correlations between stool culture, monoplex ttr, monoplex InvA, Multiplex ttr, and multiplex InvA. In this exploration, we considered all test results, whether positive or negative. To account for both censored observations and the longitudinal nature of the data, we calculated repeated measures of correlation coefficients 35 using the ranks of observations for each test (akin to a repeated-measures Spearman correlation coefficient) for measuring the correlation between the Ct values for the four molecular tests and point bi-serial correlation coefficients based on ranks for measuring correlations between standard stool culture and each of the qPCR tests ( Figure 2A). The correlation coefficients vary quite widely from 0.12 (monoplex InvA and TAC-InvA) to 0.8 (stool culture and TAC-ttr). Given that for truly negative samples, the Ct values are effectively randomly distributed near the threshold used to discriminate between positive and negative samples and that most samples were negative in most tests, the somewhat weak correlations we observe can be driven by the random Ct values for negative samples. For this reason, using only the binary negative / positive outcomes for each test, we computed    (Figure 2A). Concordance coefficients for positive ( Figure 2B) and negative ( Figure 2C) diagnosis obtained using binary negative or positive outcomes for each test. For example, in Figure 2B, the intersection of the row labeled 'Culture' and the column labeled 'ttr' lists the proportion of positive test results for the Culture test that are also positive for the ttr test. Both the size and colour depth represent the magnitude of correlation.
positive ( Figure 2B) and negative ( Figure 2C) concordance: for example, in Figure 2B, the intersection of the row labeled 'ttr,' and the column labeled 'InvA' lists the proportion of positive test results for the ttr test that are also positive for the InvA test. Unexpectedly (given that most samples were negative), negative concordance ( Figure 2C) was very high, with the lowest negative concordance being 89%. Results for positive concordance ( Figure 2B) are also relatively high, though there is more variation, ranging from 25% (for positive InvA results confirmed by positive stool cultures) to 100% (positive stool cultures confirmed by positive monoplex ttr or positive monoplex InvA).  Table 3 and Figure 1B. While stool culture has the highest PPV (99.73%) and mono-ttr the largest NPV (99.67%), a good trade-off between PPV and NPV is achieved by TAC-ttr (both high PPV, 85.97%, and high NPV, 90.33%). We note that these PPV and NPV estimates are for an asymptomatic population of children in urban Blantyre, Malawi, and are not directly generalisable to different contexts and populations where prevalence may be higher or lower.

Discussion
The burden of asymptomatic gastrointestinal exposure to Salmonella, which could be linked to either the development of immunity or, conversely, to bloodstream infection, is not known due to lack of robust Salmonella detection methods for stool specimens. This study aimed to optimise detection methods and to validate and compare the performance of monoplex ttr and InvA qPCR assays (ttr and InvA) against ttr and InvA qPCR assays on a validated multiplex qPCR platform (TACttr and TAC-InvA), and compare all molecular methods to standard Salmonella stool culture. Validation of the monoplex ttr and InvA primers showed that the primers do not cross-react with other enteric pathogens, and LOD testing showed that selenite pre-culture promotes molecular detection, even when culture is negative. Stool culture demonstrated the highest specificity but low sensitivity than all the molecular tests. Stool culture, despite having low sensitivity, still remains important in Salmonella diagnosis. Culture allows for antimicrobial susceptibility testing and strain typing. Ttr detected on the monoplex platform demonstrated superior sensitivity to stool culture, InvA, TAC-ttr, and TAC-InvA. All the test methods, however, displayed high concordance to each other.
Several studies have developed Salmonella detection methods based on antigen detection or nucleic acid amplification 16,18,36,37 . Both monoplex and multiplex nucleic acid amplification-based detection methods have been developed [38][39][40][41] . Most of these have, however, focused on Salmonella detection in blood as opposed to stool specimens. Some multiplex qPCRs to specifically detect Salmonella and its serovars or for the detection of multiple enteric pathogens in stool specimen (including Salmonella) have recently been developed 28,42-44 . The advantage of multiplex qPCR is that it is fast in determining the primary etiological agent in cases where multiple pathogens or different serovars cause the outcome, but it is expensive if one is interested in detecting only one particular pathogen (Table 4). By contrast, the advantage of a monoplex test is that it is economical than the multiplex, faster than stool culture, even with the addition of the selenite enrichment step and allows for batch-processing of samples which increases the efficiency of the test method. In this study, the same primer/ probe sets were tested using both the monoplex and multiplex qPCR platforms. The monoplex qPCR maximised sensitivity, while the multiplex panel provided a balanced pay-off between sensitivity and specificity ( Table 4). The high sensitivities of the monoplex qPCR could be attributed to the use of selenite pre-cultured stool as opposed to extraction of DNA from neat stool samples, which is used in the multiplex qPCR. Selenite sub-cultured stool samples were not used on the multiplex platform because the manufacturer's protocol was followed. Other studies have, however, also demonstrated superior performance of monoplex qPCR when compared with multiplex qPCR. The monoplex qPCR is therefore ideal for studies that are only interested in determining the presence or absence of Salmonella while capitalising on the sensitivity of the test, while multiplex qPCR will have an added advantage if a study wants to detect multiple pathogens while having a pay-off between sensitivity and specificity.
The ttr primer/ probe set used in the monoplex qPCR was previously validated for use in food samples and required validation in stool specimens. Our in-house developed InvA primer/ probe set also needed validation. Both assays demonstrated that they could detect all the different Salmonella strains, including S. Enteritidis, S. Typhimurium, and S. Given the lack of a true gold standard diagnostic test, we took a model-based approach and used an LMM to estimate the specificities and sensitivities of the 5 Salmonella detection methods. Stool culture demonstrated the highest specificity but had the lowest sensitivity. All molecular assays; TAC-ttr, of the test, specialist equipment cost by ranking them from Low to very high. Comparisons were also made for follow-up tests, antibiotic sensitivity and serotyping, and efficiency in batching. TAC-InvA, ttr, and InvA, demonstrated high specificity and sensitivity rates. Compared to the other methods, the monoplex based qPCR ttr achieved, in our opinion, the best sensitivityspecificity trade-off as it demonstrates near-perfect sensitivity (99.53%) and still achieves high specificity (95.43%). While monoplex ttr has the best overall performance, depending on the research context or clinical purpose, practitioners may still prefer tests with higher specificity such as stool culture, TAC-ttr, or TAC-InvA. All molecular test methods had significantly higher sensitivities than stool culture. High specificity and low sensitivity rates for culture have been widely reported 18 . Such low sensitivity rates should be taken into consideration when evaluating diagnostic tests. It is clear that a reference test with poor sensitivity is not adequate to evaluate alternative test methods. In such a situation, alternative means of evaluating the assays should be used, such as the LMM that has been used here. LMMs, and their counterpart for cross-sectional data, latent class models (LCMs), have been used to evaluate diagnostic tests for various pathogens, including Salmonella 48 .

Multiplex -qPCR
PPV and NPV vary depending on the prevalence of the condition being tested in any particular population. Our samples were collected from a population that was considered healthy and asymptomatic at the time of recruitment. Using the modelestimated stationary probability of being infected, we estimated the Salmonella infection prevalence of 5.25% in this population. With this prevalence estimate, stool culture demonstrated a high PPV (99.73%) when compared to molecular tests that had high NPVs, with the highest NPV recorded by mono-ttr (99.67%). As a trade-off between PPV and NPV, in a population of asymptomatic children in urban Blantyre, TAC-ttr achieved high PPV (85.97%) and high NPV (90.33%). Whether a test with high PPV or NPV is to be preferred depends on the research and/or clinical context. When prevalence is low, a slight change in specificity will have significant effects on the PPV. Higher PPVs could be observed in a situation where prevalence is high such as when using a cohort of hospitalised diarrheal cases or during a diarrhoeal outbreak.
Molecular methods had higher sensitivity but lower specificity relative to stool culture. The loss in specificity is slight compared to the gain in sensitivity, and in the case of Salmonella, the public health cost of false-negative results could be higher if the infection becomes potentially life-threatening due to withholding or delay of treatment. With the high sensitivity, molecular methods were able to detect asymptomatic Salmonella events, critical for the research questions we hoped to pose in this cohort. All the events that were detected here were asymptomatic in healthy children, which are potentially very important in transmission or the development of immunity. The detection of low bacterial burden events could also be relevant in settings like Malawi where unprescribed over-thecounter antibiotic procurement and use is common. Studies that have reported on risk factors of having a culturenegative result have indicated that antibiotic usage before sample collection is the main risk factor. Using molecular techniques such as PCR could overcome this challenge because it detects bacterial DNA regardless of pathogen viability. This might increase the probability of identifying the infection and reduce sample processing time, leading to proper patient management and treatment if needed.
Our study has several limitations. One main limitation is the use of different sample types for the two qPCR platforms. The use of selenite sub-cultured stool samples in monoplex qPCR may have contributed to the superior performance when compared with the multiplex qPCR. We used neat stool samples for multiplex qPCR to comply with the manufacturer's protocol.
Other studies have, however, demonstrated that testing primer/ probe sets in the monoplex platform perform better than in the multiplex qPCR platform. Clinical samples used to test the performance of the test are a limitation, especially in determining the PPV and NPV. Clinical samples used in the study were collected from a cohort of children that were asymptomatic to Salmonella and remained healthy for most of the one-year study period. Using samples from participants with a clinical diagnosis of Salmonella or diarrhoea would have resulted in PPV and NPV estimates more directly relevant in a clinical setting. Another important limitation is the model-based nature of our approach. This was a necessity given the lack of a gold standard diagnostic test but does mean that our estimates depend on the validity of the model's assumptions, in particular, i) the assumption that the latent infection state at a given time only depends on the immediate previous timepoint, the socalled Markov assumption, and ii) the conditional independence assumption of the basic LMM. While our modelling framework had been extended to relax the latter assumption, a better fit was achieved for the basic model.

Conclusion
The data presented here demonstrate that the addition of selenite F broth pre-enrichment step increases Salmonella detection in stool samples and that ttr and InvA primer and probe sets used can detect different Salmonella strains. The ability of ttr to detect Salmonella with such high levels of specificity and sensitivity when tested using clinical samples collected from a mostly healthy cohort makes it a promising assay that could be used for research surveillance studies. The assays could be very useful in studying the transmission of Salmonella infections. This method may perform with different sensitivity and specificity in a chronic carriage, diarrhoeal or invasive Salmonella disease state, since the load and culturability of the pathogen within the stool may be different, and further validation studies would be needed We established that selenite pre-culture increased diagnostic yield for molecular detection and identified ttr primers as molecular tools that could best help reveal the true extent of Salmonella exposure events within the gastrointestinal tract. This will allow us to understand their importance to diarrhoeal and invasive disease pathogenesis and epidemiology in the future.

1.
When using the mon-TTR and mono-InvA, the genomic DNA used is from the selenite broth cultured samples, which is time-consuming for cultivation of the stool samples, so what is the advantage of these methods? 2.
The specificity of the molecular methods used only 9 Salmonella serotypes, why not use some frequently reported human nontyphoidal Salmonella serotypes like S. London, S. Rissen, and S. Meleargidis, S. senftenberg, etc.? In addition, it is better to give the background of the prevalence of TTR and InvA in Salmonella serotypes based on genome sequencing analysis. The TAC is also used multiplex PCR methods, there is no detailed information on TAC in the manuscript.

3.
The PPV and NPV have also been used to evaluate the molecular detection methods, under the condition of an estimated 5.25% of Salmonella in asymptomatic people, which method is better based on analysis of PPV and NPV?

4.
It is necessary to revise the English grammar of the manuscript. 5.

Is the rationale for developing the new method (or application) clearly explained?
Partly

Is the description of the method technically sound? Partly
Are sufficient details provided to allow replication of the method development and its use by others? Partly If any results are presented, are all the source data underlying the results available to ensure full reproducibility? Partly Are the conclusions about the method and its performance adequately supported by the findings presented in the article? Partly content as detailed in the below responses.

Reviewer comment 1:
Application of these methods into 421 stool samples, 40, 29, 56, and 47 samples were detected positive using mono-TTR, mono-InvA, TAC-TTR, and TAC-InvA, respectively. While 23 samples were detected positive using the stool culture method. The study concluded the detection methods using TTR primers could be used for research surveillance studies but didn't confirm which method is reliable and applicable? And why? Author's response 1: We thank the reviewer for highlighting that the main conclusion was not stated clearly enough. Mono-TTR is the method which, based on our data and analysis approach, achieves, in our opinion, the best trade-off between sensitivity and specificity. However, depending on context and purpose, other researchers and practitioners may prefer to prioritise sensitivity or specificity differently and therefore prefer one of the other methods we discussed. We have clarified this in the Results and Discussion sections. In the Results section, we now state: "While stool culture achieved the highest specificity and monoplex TTR the highest sensitivity, monoplex TTR achieved arguably the best trade-off between sensitivity-and specificity: very high sensitivity (99.53%) at a relatively small drop in specificity (95.46) compared to stool culture." In the Discussion section, we now clarify: "Compared to the other methods, the monoplex based qPCR TTR achieved, in our opinion, the best sensitivity-specificity trade-off as it demonstrates near-perfect sensitivity (99.53%) and still achieves high specificity (95.43%). While monoplex TTR has the best overall performance, depending on the research context or clinical purpose, practitioners in different contexts may still choose tests with higher specificity such as stool culture, TAC-TTR or TAC-InvA". As per our response to reviewer 1, point 3, we also more clearly stated that our test performance characteristics have been derived using a statistical model. In Methods, we now state: "Another important limitation is the model-based nature of our approach. This was a necessity given the lack of a gold standard diagnostic test but does mean that our estimates depend on the validity of the model's assumptions, in particular i) the assumption that the latent infection state at a given time only depends on the immediate previous timepoint, the so-called Markov assumption, and ii) the conditional independence assumption of the basic LMM. While our modelling framework had been extended to relax the latter assumption, the better fit was achieved for the basic model."

Reviewer comment 2:
When using the mono-TTR and mono-InvA, the genomic DNA used is from the selenite broth cultured samples, which is time-consuming for cultivation of the stool samples, so what is the advantage of these methods? Author's response 2: A table showing the characteristics of Salmonella stool culture, Monoplex -qPCR ttr, Monoplex -qPCR InvA, Multiplex -qPCR ttr and Multiplex -qPCR InvA has been added to summarise the pros and cons of the test methods (Table 3). A table has been added to summarise the characteristics of the different salmonella detection methods used in this study in order to highlight the strengths and weaknesses of achieved by TAC-TTR achieving both high PPV and NPV-but again it depends on the purpose of the researcher / clinician whether to prioritise high PPV or high NPV or something in-between. We have clarified this in the manuscript. In Results we now state: "While stool culture has highest PPV (99.73%) and mono-TTR the largest NPV (99.67%), arguably the best PPV and NPV trade-off is achieved by 85.97%,and high NPV,90.33%). We note that these PPV and NPV estimates are for an asymptomatic population of children in urban Blantyre, Malawi and are not directly generalisable to different contexts and populations where prevalence may be higher or lower." And in the Discussion section we now state: "PPV and NPV vary depending on the prevalence of the condition being tested in any particular population. Our samples were collected from a population that was considered healthy and asymptomatic at the time of recruitment. When calculating the sensitivity and specificity of a method, it is necessary to determine whether the tested object is true positive or true negative. Whether the specimen is really positive can be determined by test. The author's conclusion was based on statistical analysis rather than experimental results, which I don't think is ideal.

Minor:
Introduction: Gene names should be in italics and lowercase.

Methods:
Salmonella stool culture: The volume of selenite F broth should be given. Did this method follow a literature or is it used in the authors' laboratory? Inoculating the centrifuged pellet onto XLD, I think there are too many bacteria, which may affect the subsequent selection of suspected colonies.
○ Table 1: The fluorescent groups and quenching groups of the probes needed to be given.

○
Statistical analysis: I don't understand the method used by the authors, so I don't comment on it. I suggest that the authors should explain the reliability of this method.

○ ○
Results and discussion. The analysis of cohort data is more meaningful. The reasons for the inconsistent results should be analyzed. I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.
principled approach. The cohort data allows us to detect Salmonella exposure events as they happened -useful given that Salmonella prevalence can vary across the year. If we have misunderstood the reviewer's comment, we are happy to further elaborate on the nature of the data and our analysis approach.

Reviewer comment 3:
When calculating the sensitivity and specificity of a method, it is necessary to determine whether the tested object is true positive or true negative. Whether the specimen is really positive can be determined by test. The author's conclusion was based on statistical analysis rather than experimental results, which I don't think is ideal.

Authors' response 3:
We thank the reviewer for highlighting that we were not clear enough about how we estimated the test performances. The reviewer is right to point out that we used a statistical model for assessing the various diagnostic methods. If a true, perfect gold standard method existed, then it would, of course be preferable to compare the methods against this perfect diagnostic test. However, in our case the only available reference standard, culture, is known to have low sensitivity. For this reason, we decided to use the statistical model described in the manuscript and also, in more detail, in a separate publication (Henrion, 2018) The statistical model we used is of a particular type, specifically it is a latent Markov model (LMM), an accepted approach to model the underlying true infection state of individuals. The model uses this modelled latent infection state to estimate the sensitivity and specificity parameters that are most likely to have given rise to the observed data. As with any statistical model, there are a number of assumptions that have been made (in particular the LMM used in the present manuscript assumes i) that the infection state at a given timepoint only depends on the immediate prior timepoint and ii) that conditional on the latent state, the response probabilities of the different diagnostic tests are independent) and if these are not met, the results could be invalid. Despite the need to use a statistical rather than an experimental approach, we have done some experimental validation. using spiked stool samples to assess detection limits for the molecular methods. We could not, however, use such spiked samples to calculate sensitivities and specificities: such artificial samples do not reflect true, biological samples and could lead to estimates for sensitivity and specificity that are biased to incorrectly high values. For this reason, we preferred to use real world data and a statistical approach to best estimate the test characteristics. Please note that this is a well-recognised and validated approach in this situation, and similar models to ours have been used by other authors to evaluate novel diagnostic tests (Koukpunari A, 2013) We have made the following revisions to the manuscript to incorporate and clarify the above points: In the Discussion section where we discuss the specificities and sensitivities that we estimated, we now clarify this is a model-based approach: ○ "Given the lack of a true gold standard diagnostic test, we took a model-based approach and used an LMM to estimate the specificities and sensitivities of the 5 Salmonella detection methods." In the Discussion section, we now list the model-based approach as another limitation, given that our results depend on the validity of the model assumptions.