Variation in dengue virus plaque reduction neutralization testing: systematic review and pooled analysis

Search strategy and selection criteria

A systematic review was conducted during May-June 2009 using the PubMed and ISI Web of Knowledge 4.0 databases. Nine searches were performed using the search term “dengue” and each of the following terms or phrases: “antibody”, “serology”, “neutralization”, “serum neutralization”, “prnt”, “infection history”, “previous infection”, “longitudinal”, “antigenicity”, “vaccine”, “chimerivax” and “over time”. Articles were not restricted by year of publication and included electronic, ahead-of-print publications available in these databases. We did not attempt to contact authors to obtain unpublished PRNT titers. Abstracts were screened by two individuals (KRL and IRB). Articles in languages other than English, Spanish, and Portuguese and case reports or studies of unusual patient populations were excluded. Unusual patient populations were defined as individuals with co-morbidities (e.g., cancer or organ failure) or neurologic or central nervous system manifestations. Articles that did not report dengue PRNT titers from humans or did not state the time from dengue virus exposure to sera collection were also excluded.

Assessment

Each article was categorized based on the type of sera samples collected. Natural exposure studies included serological surveys (sera collected from individuals regardless of dengue-like symptoms), infection studies (sera collected in response to suspected dengue infection as determined by study investigators or clinic/hospital staff), and human inoculation studies or accidental inoculations in a laboratory setting. Vaccine studies were defined as the collection of sera samples after enrollment in a DENV vaccine study.

Data abstraction

Titers against each of the four DENV serotypes were recorded from individual human subjects and classified as natural exposure or vaccination. Titers reported in tables or the text were abstracted while titers illustrated in figures or requiring calculations by abstractors (e.g., an index score) were excluded. For the descriptive analysis, minimum titers (e.g., < 1:10 or < 1:40) were set to zero and end-point titers that exceeded the maximum dilution specified in the article, such as ≥ 1:2560 or ≥ 1:5120, were set to that value. The recorded materials and methods used in PRNTs for each article included dengue serotypes and strains, virus concentrations, cell lines, and neutralization percentages (required percentage of plaques neutralized). If possible, authors were contacted via email to obtain information on the DENV strains used for PRNT; however, other raw data were not requested.

The approximate time from infection or vaccination to serum sample collection was categorized as unexposed, acute (< 12 days), convalescent (12–30 days), late convalescent (30 days–1 year), and very late convalescent (> 1 year) [13, 14]. If a range of time was reported for an individual sera sample (e.g., “Serum was collected 6–12 months following infection…”), the mid-point between the minimum and maximum of the range was used. If acute and convalescent sera samples were collected but exact times from exposure or symptom onset were not reported, acute samples were defined as 4 days post-exposure and convalescent samples defined as 11 days post-exposure. If time from symptom onset to serum sample collection was reported, 5 days were added to account for the 4–6 day incubation period of dengue virus [13] and make times since exposure comparable between post-symptom onset and post-exposure reports.

Previous DENV exposure or infection (primary/secondary/tertiary) was based on explicit report by the authors. “Confirmed” exposure to a specific serotype was based on the reported contents of a vaccine in a vaccine study or inoculation suspension in inoculation studies or a polymerase chain reaction (PCR)-confirmed infection in natural exposure studies. If these criteria were not met, the probable infecting serotype was noted by abstractors but not considered confirmed. For example, probable infecting serotypes were commonly deduced by an article’s authors from changes in serotype-specific antibody titers or the most frequently detected serotype during a recent epidemic.

Pooled analysis

Pooled analyses were conducted using PRNT titers for which a confirmed exposure serotype and exposure history (e.g., uninfected, primary, or secondary) could be determined. These analyses aimed to characterize the effects of strain choice and inter-laboratory variation on PRNT titers after adjusting for laboratory methods, exposure route and time since exposure. In the case of primary exposure, it was assumed that natural logarithm-transformed (log) titers depended on the exposure serotype. For secondary exposure it was assumed that log titers were independent of the exposure serotype, as indicated by the descriptive analysis of this data and prior studies [15]. Pooled median log titers for each exposure-testing serotype were calculated as a weighted mean of strain-specific medians with the number of strain-specific observations as weights.

We constructed log-linear hierarchical models to characterize the effects of strain choice and inter-laboratory variation on titers. Titers from individuals who were simultaneously exposed to two or more serotypes via multitypic vaccination were excluded. Titers reported as less than the minimum dilution for that particular article (e.g., <1:10 or <1:40) were treated as censored observations distributed between 0.1 and the minimum reported titer in that article. The maximum reported titers were also treated as censored observations distributed between the maximum reported titer and a 100-fold increase.

Model A assessed the effect of strain choice on log PRNT titer. Categorical strain variables were assumed to have random effects with mean zero and estimated variance. Strains originating in Thailand were used as the reference strain within each serotype due to frequent use in the abstracted articles and to provide geographic consistency. Strains reported by only one article were categorized as “Other” in each serotype. The strain categories were as follows: DENV1 – 16007, Hawaii, Other; DENV2 – 16681, New Guinea B, New Guinea C, PR-159, Other; DENV3 – 16562, H87, 116/00, Other; DENV4 – 1036, H241, Dominica /814669, 4328S, Other. Model B assessed the effect of inter-laboratory variation on log PRNT titer. A random effect for article, with zero mean and estimated variance, was included in this model. Due to collinearities between the strain choice and article, a model with effects for both strain and article was not evaluated.

In both models fixed effects were estimated to account for infecting serotype and other factors that might affect measured PRNT titer. For primary exposures, intercepts were estimated for each exposure serotype-testing serotype pair while for secondary exposures, intercepts were estimated for each testing serotype. The following were also evaluated during model selection: time since exposure, neutralization percentage, cell line, virus concentration and vaccination. Neutralization percentage estimates the effects of each 10% increase in neutralization percentage after adjusting for other covariates. Similarly, the effect of vaccination is estimated with reference to individuals with natural exposure after adjusting for other covariates. Time since exposure was treated as an ordinal variable using categories described above with unexposed sera as the reference category.

All models were fit using Monte Carlo Markov Chain (MCMC) methods in OpenBUGS [16] and the R Statistical Package (version 2.10, R Development Core Team, Vienna, Austria). Standard non-informative priors were used for all fixed effects and estimated variance components. For each model, three chains of 10,000 iterations were run. The final 5,000 iterations of all chains were combined to estimate the posterior distribution of parameters. Models were selected based on minimizing the deviance information criteria (DIC). The mean of the posterior distribution is reported as the parameter estimate, and 90% credible intervals report the range between the 5^th and 95^th percentiles of the posterior distribution. Convergence was assessed by examination of the MCMC chains and R-hat values for all estimated parameters of <1.1 [17]. Full specifications of the models are in Additional File 1.

Variation in laboratory methods

Twenty-nine articles (90%) reported DENV1 PRNT titers, nine (28%) of which did not report the strains used (Table 2). The Hawaii, 16007, and Mochizuki strains were used in multiple articles and six additional strains were reported in one article each (indicated by “Other” in Tables 2 and 3). Of these 29 articles, three (10%) tested multiple DENV1 strains. Three articles did not test against DENV1 strains (indicated by “n/a”). Seventeen strains of DENV2 were tested in thirty articles and two articles (7%) reported PRNT titers against multiple strains. Titers against DENV2 strain 16681 were reported most frequently (n = 7 articles) followed by New Guinea C (n = 4), New Guinea B (n = 3), and PR-159 (n = 2). Two articles did not report PRNT titers against DENV2 strains. Titers against DENV3 strains were reported in 31 articles with strain H87 used in nine articles (28%). Seventeen different DENV3 strains were reported with two articles (6%) testing multiple DENV3 strains. Eight DENV4 strains were used in 27 studies and one article reported testing multiple strains. H241 was used most frequently (n = 9 articles), followed by 4328-S (n = 3) and Dominica/814669 and 1036 reported in 2 articles each.

The cell lines, neutralization percentages, and concentration of virus added to the PRNT were highly variable between articles (Table 1). Plaque-forming units (pfu) and focus-forming units (ffu) (quantities of viral particles differentiated by infection of neighboring cells in an assay) varied widely and were combined with varying volumes of sera. LLC-MK₂ cells were used in fifteen articles (47%) while six and nine articles (19% and 28%) reported using Vero or BHK-21 cells, respectively. Two articles (6%) did not report the cell lines used and one article reported using both BHK-21 and LLC-MK₂ cells for the purposes of comparison [31]. Twenty-three articles reported neutralization percentages of 50%, while four articles reported 70%, two articles reported 80%, and one article reported 90%. One article compared neutralization percentages of 50% and 70% [31] and three articles did not report the neutralization percentage used [27, 30, 43].

Patterns of reported PRNT titers in primary and secondary exposure

Titers were excluded from the pooled analysis if the exposure serotype was not reported (n = 786), if exposure occurred via multitypic vaccine suspension (n = 944) or if titers were from unexposed sera (n = 74), leaving 1,615 primary and secondary exposure titers. Among 987 titers from primary exposures, titers to the exposure serotype were consistently high, and there was low reactivity to non-exposure serotypes (Figure 3). When sera from confirmed DENV1 exposures were tested against DENV1 strains, the pooled median log titer was 4.75 (standard error [SE]: 0.049) while testing DENV1-exposed sera with DENV2, DENV3, and DENV4 strains produced pooled median log titers ranging from 0.14-1.51. Homotypic testing with DENV2, DENV3, and DENV4 resulted in pooled median log titers of 4.98 (SE: 0.103), 4.45 (SE: 0.119) and 5.49 (SE: 0.118), respectively, while heterotypic testing produced titers ranging from 0–0.288.

In contrast, secondary DENV exposures (n = 628 observations) produced high heterotypic titers, and testing with the exposure serotype did not always generate the highest PRNT titer (Figure 4). Pooled median log titers were highest when tested against the exposure serotype for DENV1 (6.89 [SE: 0.094]) and DENV2 (7.36 [SE: 0.042]) compared to non-exposure serotypes. For DENV3 exposure, however, the pooled median log titer from DENV3 testing (3.98 [SE: 0.059]) was lower than titers produced by DENV1 testing (5.63 [SE: 0.039]). Secondary DENV4 exposures showed a similar trend of high pooled median log titers against cross-serotypes, but few observations (n = 16) yielded uninformative statistical comparisons. As a result of these observations, the PRNT titer resulting from testing after secondary exposure was assumed to be independent of the secondary exposure serotype in the hierarchical models.

Strain dependence of reported PRNT titers

Among individuals with confirmed primary DENV exposures, neutralization titers from each strain varied substantially within each serotype (Figure 3). For instance, individuals with primary DENV3 exposure had variability in unadjusted median log titers between strains ranging from 2.89 (IQR: 2.30, 3.23) for strain CH53489 to 5.56 (IQR: 5.21, 6.75) for strain 16562. Median log titers among individuals with secondary exposure also showed variability between strains; however, stratification by strain resulted in few observations and limited inference (Figure 4).

In the hierarchical models, strain variation accounted for 8.04% (90% CrI: 3.05%, 15.7%) of the variability in the PRNT titers after adjusting for other covariates (Table 4, Model A). While most strains produced relatively small effects on titers in comparison to the reference strains, others yielded more significant changes. Testing with DENV2 strain PR-159 and DENV4 strain Dominica/814669 yielded titers that were 0.48 [90% CrI: 0.25, 0.85] and 0.32 [90% CrI: 0.16, 0.62] that of titers from the respective reference strains, while DENV4 strain 4382-S produced 1.62-fold (90% CrI: 1.02, 2.55) increase in titer compared to the reference strain (Table 4, Model A). Additionally, accounting for strain in Model A reduced the effect of secondary exposure by 36-79% (Model A vs. B in Table 4).

Other drivers of variation in PRNT titers

Time since exposure demonstrated a characteristic pattern of serologic response in both models, although we were unable to estimate the effect of acute DENV exposures due to a lack of titers fulfilling model inclusion criteria (Table 4). In Model A, convalescent DENV exposures (n = 339 titers) produced 8.51-fold (90% CrI: 5.47, 13.2) the mean titer of unexposed individuals (n = 177), decreasing to 5.57-fold (90% CrI: 3.79, 8.24) during late convalescence (n = 515) and 3.13-fold (90% CrI: 2.05, 4.81) during very late convalescence (n = 608). Each 10% change in neutralization percentage resulted in a 1.23-fold (90% CrI: 1.14, 1.33) change in titer (Table 4, Model A). Interestingly, titers resulting from vaccination (n = 433) were 0.91 (90% CrI: 0.70, 1.16) that of titers from natural exposure (Table 4). Cell line and virus concentration were evaluated as potential covariates but did not explain additional variation among titers as determined by increases in DIC values (data not shown).

Variation between articles (a proxy for inter-laboratory differences) produced 50.7% (90% CrI 30.8%, 71.6%) of the variation in PRNT titers not accounted for by other covariates (Table 4, Model B). Because of the association between article and laboratory methods, it is not possible to quantify how much of this variation is due to differences in strain and other aspects of the PRNT methods while also accounting for article variation. It is important to note that article-to-article variation may be inflated by differences in study population and other unmeasured covariates.