Modelling the immunological response to a tetravalent dengue vaccine from multiple phase-2 trials in Latin America and South East Asia

Background The most advanced dengue vaccine candidate is a live-attenuated recombinant vaccine containing the four dengue viruses on the yellow fever vaccine backbone (CYD-TDV) developed by Sanofi Pasteur. Several analyses have been published on the safety and immunogenicity of the CYD-TDV vaccine from single trials but none modelled the heterogeneity observed in the antibody responses elicited by the vaccine. Methods We analyse the immunogenicity data collected in five phase-2 trials of the CYD-TDV vaccine. We provide a descriptive analysis of the aggregated datasets and fit the observed post-vaccination PRNT50 titres against the four dengue (DENV) serotypes using multivariate regression models. Results We find that the responses to CYD-TDV are principally predicted by the baseline immunological status against DENV, but the trial is also a significant predictor. We find that the CYD-TDV vaccine generates similar titres against all serotypes following the third dose, though DENV4 is immunodominant after the first dose. Conclusions This study contributes to a better understanding of the immunological responses elicited by CYD-TDV. The recent availability of phase-3 data is a unique opportunity to further investigate the immunogenicity and efficacy of the CYD-TDV vaccine, especially in subjects with different levels of pre-existing immunity against DENV. Modelling multiple immunological outcomes with a single multivariate model offers advantages over traditional approaches, capturing correlations between response variables, and the statistical method adopted in this study can be applied to a variety of infections with interacting strains.


Data
In the descriptive analysis we selected all subjects in the vaccine group with complete records of baseline, post-dose 1 (PD1), post-dose 2 (PD2) and post-dose 3 (PD3) titres against the four dengue serotypes and with baseline titres against JEV (see Table 1 in the main text). In the multivariate regression analysis we included all subjects in the vaccine group with complete records of baseline and PD3 titres against all four dengue serotypes and baseline JEV titre. Table S1 describes the baseline demographic and immunological characteristics of the subjects included in the multivariate regression analysis.  : Baseline demographic characteristics of the subjects included in the multivariate regression analysis. Study denotes the trial identifier; Location denotes the location where the trial was conducted, i.e. the country if the trial was single-site or the region if the trial was multicentre; Age denotes the observed mean and 5-95 percentiles of the observed age of the subjects included in the analysis in years; N denotes the number of subjects; JEV+ denotes subjects with titre against JEV ≥ 10; DENVdenotes subjects with titres < 10 for all four DENV serotypes; DENV+ denotes subjects with titre ≥ 10 for at least one DENV serotype; monotypic denotes subjects with titre ≥ 10 for one DENV serotype or titre ≥ 10 to more than one serotype with titre ≥ 80 to one DENV serotype only; multitypic denotes subjects with titre ≥ 10 for more than one DENV serotype without titre ≥ 80 to only one serotype. The percentages within parentheses are computed on the number of subjects in each study N. All titres have been quantified using PRNT50.

Notation
We denote Y ji the log10 PD3 titre of subject j against DENV i and W jk the set of predictors for subject j, k = 1, . . . , p. The covariates used in the regression models contain immunological (baseline titres against DENV and JEV and immunological status against DENV and JEV), demographic (gender, age) and geographical (trial code, trial location) information for each subject in the study. The covariates and the notation used are described in Table S2.
T 5 j dummy 1 if subject j is part of trial T5, 0 otherwise A 2j dummy 1 if subject j is ≥ 5 and ≤ 10 years old at baseline, 0 otherwise A 3j dummy 1 if subject j is > 10 years old at baseline, 0 otherwise Table S2: Summary of variables adopted in the analysis.

Assessing the joint normality of the data
We judge the joint normality of the data through visual inspection of the chi-squared plot of the data.
Denote Y j· the sample observations for subject j,Ȳ i the sample mean of the PD3 titre against DENV i and Y the vector of samples means against DENV1-4. Let S denote the sample variance-covariance matrix of the observed PD3 titres. The squared generalized distances are defined where the superscript t denotes the transpose operation and the superscript −1 denotes the inverse operation.
The chi-squared plot in Figure S1 is obtained by plotting the pairs ( d 2 (j) , χ 2 where d 2 (j) is the jth ordered squared distance and χ 2 p ( j−1/2 N ) is the 100 j−1/2 N percentile of the chi-square distribution with 4 degrees of freedom. Visual inspection of the graph in Figure S1 shows that we can assume that the distribution of the PD3 titres is multivariate normal.

Model variants analysed
In the main text (eq. (1)) we presented the general model formulation. In this section we describe all model variants explored in our analysis. The models do not depend on the value assigned to the undetected titres, since we only included the titres (or summary statistic of the titres) when these are quantified (i.e. ≥ 10) and estimate the relative contribution of the undetected titres (or summary statistics involving undetected titres) when the titres involved are below the detection limit.

Modelling the PD3 DENV titres using the baseline DENV titres as covariates
The models presented in this section (Models 1-11) use the baseline DENV titres as predictors of the PD3 DENV titres. Model 1 models the PD3 titre of DENVi Y ji using the homologous baseline titre X ji as covariate, if the homologous baseline titre is detected (i.e. if U ji = 0): The covariates used in Model 2 are the homologous baseline titre (as done in Model 1) and the maximum of the heterologous baseline titre, if detected (i.e. if Z ji = 0): The covariates used in Model 3 are the homologous baseline titre (as done in Model 1) and the average of the heterologous titres, if all heterologous titres are detected (i.e. if V ji = 1): Model 4 includes all four baseline titres as covariates. In Model 4 the coefficient of the homologous baseline titre is serotype-specific (β 1i ) whilst the coefficients of the heterologous baseline titres do not depend on the response's serotype (β 3k ): Model 5 includes all four baseline titres as covariates and we estimate a response's serotype-specific coefficient for each baseline titre (β ki ): Models 6-11 extend Models 2-5 by including titres interactions. Model 6 extends Model 2 and includes the interaction between the homologous baseline titre and the maximum of the heterologous baseline titres, if the interacting terms are detected (i.e. if U ji = 0 and Z ji = 0): Model 7 extends Model 3 and includes the interaction between the homologous baseline titre and the average of the heterologous baseline titres, if all titres are detected (i.e. if U ji = 0 and V ji = 1): Model 8 extends Model 4 and includes all interactions between the baseline homologous and heterologous titres, if the interacting titres are detected (i.e. if U ji = 0 and U jk = 0 for k ̸ = i): maximum of the heterologous titres, if detected (i.e. if U ji = 0 and Z ji = 0): Model 10 extends Model 5 and includes all interactions between the baseline homologous and heterologous titres, if the interacting titres are detected (i.e. if U ji = 0 and U jk = 0 for k ̸ = i): Model 11 extends Model 5 and includes the interactions between the baseline homologous titre and the maximum of the heterologous titres, if detected (i.e. if U ji = 0 and Z ji = 0): (11) Note that term U ji Z ji in equations (6) and (9) is a seronegative-specific term, i.e. U ji Z ji = 1 if and only if subject j has undetected titres to all 4 serotypes.

Building on model 9 (the best model)
We build on Model 9 defined in (9) and explore the effect of adding one individual covariate, other than the baseline DENV titres, to the regression model. We include each covariate additively, i.e. adding a serotype-specific intercept for each predictor. In Models 9a-9g we respectively include the gender, the age of the subject, the trial location (i.e. the continent) where the subject was enrolled, information on the baseline immunological DENV status of the subjects, the JEV titre (if detected), information on the baseline immunological status against JEV and the trial-specific identifiers.
The equation of model 9g, which includes serotype-specific intercepts for each trial, is Models 9h-9l build on model 9g and include an additional individual covariate to the regression mode, specifically the gender, the age of the subject, the baseline immunological DENV status of the subjects, the JEV titre (if detected), the baseline immunological status against JEV, respectively. The equation of model 9j, which includes serotype-specific intercepts for individuals with a single baseline seropositive DENV titre, is Models 9m-9p build on Model 9j and include the gender, the age of the subject, the JEV titre (if detected) and the baseline immunological status against JEV, respectively.

Building on model 11
Similarly to the analysis presented in section 1.4.2, we build on Model 11 defined in (11) and explore the effect of adding one individual covariate, other than the baseline DENV titres, to the regression model. We include the extra covariates additively, i.e. adding serotype-specific intercepts for each covariate. In Models 11a-11g we respectively include the gender, the age of the subject, the trial location where the subject was enrolled, information on the baseline immunological DENV status of the subjects, the JEV titre (if detected), information on the baseline immunological status against JEV and the trial-specific identifiers.
The equation of model 11g, which includes serotype-specific intercepts for each trial, is Models 11h-11l build on model 11g and include an additional individual covariate to the regression mode, specifically the gender, the age of the subject, the baseline immunological DENV status of the subjects, the JEV titre (if detected), the baseline immunological status against JEV, respectively. The equation of model 11j, which includes serotype-specific intercepts for individuals with a single baseline seropositive DENV titre, is East Asia (SEA) or Latin America (LA)) (row 1), by baseline immunological status against DENV and trial (row 2) and by baseline immunological status against DENV and JEV (only for South East Asian studies) (row 3). The rise r 12 between two titres T t 1 , T t 2 taken respectively at time points t 1 , t 2 is defined as the number of 2-fold serial dilutions needed to obtain T t 2 from T t 1 , i.e. r 12 = log 2 (T t 2 /T t 1 ).

Model comparison and parameter estimates
The results obtained with models 1-11 are given in Table S3 and show that model 9 has the lowest AIC among models 1-11. Model 11 shows a comparable goodness of fit to model 9 (∆AIC < 4) but model 9 is more parsimonious, with 12 fewer parameters than model 11. Table S4 shows the results, in terms of goodness of fit, obtained with model variants 9a-9p. According to the AIC, model 9j is the best model. Table S5 shows the estimates obtained with model 9j and Table S6 shows the correlation between the PD3 titres against the four DENV serotypes computed from the estimated variance-covariance terms using model 9j. The correlation ρ il between the PD3 titres against the four DENV strains is computed from the maximum likelihood estimates of the variance-covariance termsΣ il according to the definition Figure S6 shows the observed and expected PD3 titres versus the homologous baseline titre. The expected average PD3 titres are computed using the maximum likelihood estimates obtained with model 9j (see Table   S5) and the noise is simulated as a single random realisation of the multivariate normal distribution with zero mean and the estimated variance-covariance matrix Σ using model 9j.
The results obtained with models 11a-11p are given in Table S7. Model 11m has the lowest AIC overall but the improvement in terms of AIC from model 11j is small (∆AIC < 4), hence we choose the more parsimonious model as the best fitting model.  Table S3: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of parameters estimated in the regression equation (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to the specified covariates, using the regression models in equations (1)-(11). The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4). The AIC of the best model is in bold.  Table S4: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to the baseline titres as in model 9 plus the specified covariates. Models 9a-9g, 9h-9l and 9m-9p have respectively 1,2 and 3 additional covariates compared to model 9. The AIC of the best model in each section of the table is in bold and the AIC differences are relative to the best model in each section. The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).

Serotype i=1
Serotype i=2 Table S7: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to the baseline titres as in model 11 plus the specified covariates. Models 11a-11g, 11h-11l and 11m-11p have respectively 1,2 and 3 additional covariates compared to model 11. The AIC of the best model in each section of the table is in bold and the AIC differences are relative to the best model in each section. The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).

Assessing the predictive power of the best fit models
We adopt a 2-fold cross-validation approach and assess the predictive power of the best models (model variants 9j and 11j) by repeating 100 times the following procedure: 1. randomly split the observed data into training T and validation V sets of indexes of equal sizes 2. fit the model to the training dataset T 3. using the MLE obtained in step 2. computeŶ ji , the average predictions in the validation set V 4. compute a serotype-dependent R 2 i for the training and validation sets, given by where K = T or K = V andȲ i denotes the average for serotype i obtained in the training dataset T 5. swap the training and validation sets and repeat steps 2.-4.
We finally compute the average R 2 i obtained in the 100 iterations of steps 1-5 for the training and validation sets.  Table S10: Average R 2 s obtained in the 2-fold cross-validation analysis using model variants 9j and 11j.

Exploring further variants of model 9
In this section we explore the effect of omitting a seronegative-specific intercept and introducing a serotypespecific intercept for profiles with a single seropositive DENV titre in model 9, given in equation (9).
The equation of model 9 without a seronegative-specific intercept (U ji Z ji ) is The equation of model 9 with a serotype-specific intercept for the profiles with a single detected titre against the homologous DENV serotype i (Z ji ) is The equation of model 9 having omitted U ji Z ji and added Z ji is: The results obtained with the variants of model 9 given in equations (17)- (19) are presented in Table S11.
We find that the omission of a seronegative-specific intercept significantly worsens the fit of model 9, whilst the introduction of term Z ji does not significantly improve the fit of the model to the data.  Table S11: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to Models given in equations (9) and (17)- (19). The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).

With covariates on the slope
In this section we test the effect of including in model 9 each individual covariate multiplicatively, i.e. on the slope rather than on the intercept. This means that the covariates interact with the baseline titres, if detected. For example, when introducing the gender on the slope, we estimate separate coefficients of the baseline (detected) titres for males and females. We denote the models with covariates on the slope models 9S 1 -9S 5 . The covariates included in models 9S 1 -9S 5 are respectively the gender, the trial location (i.e. the continent where the trial was conducted), the age (classified into three categorical classes < 5, 5 − 10, > 10 years), the baseline immunological status against JEV and the trial-specific identifiers. Model 9S 5 , which includes the trial identifier on the slope, estimates a separate coefficients for the baseline (detected) titres for each trial. The equation of model 9S 5 is given by Table S12 summarizes the results obtained with model variants 9S 1 -9S 5 and, for convenience of comparison, also of model 9. Introducing study-specific slopes improve model 9 but the improvement is much smaller than the one obtained introducing study-specific intercepts (i.e. using model 9g, see Table S4). For this reason we did not extend the analysis with models 9S 1 -9S 5 further.  Table S12: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to Models 9 and 9S1-9S5. The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).
In this section we explore the effect of adding, to model variant 11, further information on the immunological status against DENV. Model 11 with a seronegative-specific intercept (U ji Z ji ) is given by (21) The equation of model 11 with the extra covariate Z ji , which is specific to the subjects with undetected heterologous titres (k ̸ = i), is (22) The equation of model 11 with the addition of a seronegative-specific intercept U ji Z ji and a serotype-specific intercept for the profiles with a single detected titre against the homologous DENV serotype i (Z ji ) is Table S13 summarises the results obtained with the variants of model 11 given in equations (21)-(23). The results in Table S13 show that models given in (21) Table S13: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to models given in equations (11) and (21)-(23). The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).

With covariates on the slope
Similarly to the analysis presented in section 3.1.1, here we test the effect of including in model 11 each individual covariate multiplicatively, i.e. on the slope rather than on the intercept. We denote the variants of model 11 with covariates on the slope models 11S 1 -11S 5 . The covariates included in models 11S 1 -11S 5 are respectively the gender, the trial location (i.e. the continent where the trial was conducted), the age (classified into three categorical classes < 5, 5 − 10, > 10 years), the baseline immunological status against JEV and the trial-specific identifiers. Model 11S 5 , which includes the trial identifier on the slope, estimates a separate coefficients for the baseline (detected) titres for each trial. The equation of model 11S 5 is given by (24) Table S14 summarizes the results obtained with model variants 11S 1 -11S 5 and, for convenience of comparison, also of model 11. Introducing the continent-specific identifier improves model 11 but the improvement is considerably smaller than the one obtained through the introduction of trial-specific intercepts (i.e. using model 11g, see Table S7). For this reason we did not extend the analysis using model variants 11S  Table S14: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to Models 11 and 11S1-11S5. The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).

Alternative definitions of the baseline immunological status against DENV and JEV
In this section we present the results obtained using alternative definitions of DENV seronegative, monotypic and multitypic, JEV seronegative and JEV seropositive to the ones adopted in the manuscript. The definition adopted in the manuscript, here referred as definition 1, classifies the subjects with titres < 10 to all 4 DENV serotpyes as seronegative, subjects with titre ≥ 10 to a single DENV serotype or with titre ≥ 10 to more than one DENV serotype with a titre ≥ 80 to only one serotype as monotypic and subjects with titre ≥ 10 to more than one DENV serotype without titre ≥ 80 to a single serotype as multitypic.

Definition 2 with threshold at 10
In this section we adopt the following definitions: an antibody pattern is classified as seronegative if it has titre < 10 to all 4 DENV serotpyes, monotypic if it has titre ≥ 10 to one DENV serotype only and multitypic if it has titre ≥ 10 to more than one DENV serotypes. We define a JEV titre < 10 to be seronegative (JEV-) and a JEV titre ≥ 10 to be seropositive (JEV+). Figure S7 shows the mean and 95% confidence interval (CI) of the observed DENV titres and raises in titres in time in DENV seronegative, monotypic and multitypic profiles, according to the definition given in this section. Figures S8 and S9 show the mean and 95% confidence interval (CI) of the observed DENV titres in time and of the observed rises in titres between successive vaccinations respectively, having classified the subjects as DENV seronegative, monotypic profile with detected DENV1 titre, monotypic profile with detected DENV2 titre, monotypic profile with detected DENV3 titre, monotypic profile with detected DENV4 titre or multitypic infection profile, according to the definition given in this section.

Definition 2 with threshold at 40
In this section we adopt the same definition given in section 3.3.1 but with threshold equal to 40 instead of 10. An antibody pattern is classified as seronegative if it has titre < 40 to all 4 DENV serotypes, monotypic if it has titre ≥ 40 to one DENV serotype only and multitypic if it has titre ≥ 40 to more than one DENV serotypes. For consistency, we define a JEV titre < 40 to be seronegative (JEV-) and a JEV titre ≥ 40 to be seropositive (JEV+). Figure S10 shows the mean and 95% confidence interval (CI) of the observed DENV titres and raises in titres in time in DENV seronegative, monotypic and multitypic profiles, according to the definition given in this section. Figures S11 and S12 show the mean and 95% confidence interval (CI) of the observed DENV titres in time and of the observed rises in titres between successive vaccinations respectively, having classified the subjects as DENV seronegative, monotypic profile with detected DENV1 titre, monotypic profile with detected DENV2 titre, monotypic profile with detected DENV3 titre, monotypic profile with detected DENV4 titre or multitypic infection profile, according to the definition given in this section. Figures S13 and S14 show the mean and 95% CI of the observed DENV titres in time and of the observed rises in titres between successive vaccinations stratified by baseline immunological status against DENV and continent where the trial was conducted (i.e. South East Asia (SEA) or Latin America (LA)) (row 1), by baseline immunological status against DENV and trial (row 2) and by baseline immunological status against DENV and JEV (only for South East Asian studies) (row 3). In terms of multivariate analysis, we explored the effect of using the threshold of 40 to define undetectable (< 40) and detectable (≥ 40) titres. Under this assumption, the results obtained with models 1-11 are given in Table S15 and show that model 8 has the lowest AIC.
Models 8a-8g build on Model 8 and include a serotype-specific intercept for the gender, the age of the subject, the continent where the subject was enrolled, the baseline immunological DENV status of the subjects, the JEV titre (if detected), the baseline immunological status against JEV and the trial, respectively. The equation of Model 8g, which includes serotype-specific intercepts for each trial, is (25) Models 8h-8l extend Model 8g and add a serotype-specific intercept for the gender, the age of the subject, the baseline immunological DENV status of the subjects, the JEV titre (if detected) and the baseline immunological status against JEV, respectively. Table S16 shows the results, in terms of goodness of fit, obtained with model variants 8a-8l. According to the AIC, model 8g is the best model overall. Tables S17 and S18 show the estimates and the correlation between the PD3 titres against the four DENV serotypes obtained using model 8g, respectively. Figure S15 shows the observed and expected mean and 5-95 percentiles of the PD3 titres versus the homologous baseline titres obtained with the best model (Model 8g), having simulated 100 independent realisations of the error by drawing from the multivariate normal distribution with zero mean and the estimated variance-covariance matrix Σ given in Table S17. Figure S16 shows the scatterplot of the observed and expected PD3 titres versus the homologous baseline titres having simulated a single random realisation of the error.  Table S15: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of parameters estimated in the regression equation (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to the specified covariates, using the regression models in equations (1)-(11). The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4). The AIC of the best model is in bold.  Table S16: Akaike's Information Criterion score (AIC), AIC difference from best model (∆AIC), log-likelihood (Log-like), number of estimated parameters (np) and fraction of serotype-specific explained variance compared to the null model (R 2 1 , . . . , R 2 4 ) obtained by fitting the observed post dose 3 titres to the baseline titres as in model 8 plus the specified covariates. Models 8a-8g and 8h-8l have respectively 1 and 2 additional covariates compared to model 8. The AIC of the best model in each section of the table is in bold and the AIC differences are relative to the best model in each section. The number of estimated parameters (np) includes the coefficients of the regression models β ki (k = 0, . . . , p) and the variance-covariances of the serotypes Σ il (i, l = 1, . . . , 4).      Figure S16: Observed (red), average expected (black) and expected with simulated noise (green) PD3 titres versus the observed homologous baseline titre, obtained with model 8g. The noise has been simulated by drawing a single realisation from a multivariate normal distribution with zero mean and the variance-covariance matrix Σ estimated with model 8g. Titres are shown on a log10 scale.

Definition 3 with threshold at 10
In this section we present the results of the descriptive analysis obtained using the following definitions: an antibody pattern is classified as seronegative if it has titre < 10 to all 4 DENV serotpyes; monotypic if it has titre ≥ 10 to one DENV serotype only or if it has titre ≥ 10 to more than one DENV serotypes with ≥ 6 fold higher response of the dominant (i.e. highest) DENV serotype versus the second dominant DENV serotype; multitypic if it has titre ≥ 10 to more than one DENV serotypes with < 6 fold higher response of the dominant (i.e. highest) DENV serotype versus the second dominant DENV serotype. For monotypic subjects, the dominant DENV serotype is assumed to be the serotype of the previous DENV infection. Figure S17 shows the mean and 95% confidence interval (CI) of the observed DENV titres and raises in titres in time in DENV seronegative, monotypic and multitypic profiles, according to the definitions given in this section. Figures S18 and S19 show the mean and 95% confidence interval (CI) of the observed DENV titres in time and of the observed rises in titres between successive vaccinations respectively, having classified the subjects as DENV seronegative, monotypic profile with detected DENV1 titre, monotypic profile with detected DENV2 titre, monotypic profile with detected DENV3 titre, monotypic profile with detected DENV4 titre or multitypic infection profile, according to the definitions given in this section.

Definition 3 with threshold at 40
In this section we adopt the same definitions given in 3.3.3 but with threshold at 40 instead of 10. An antibody pattern is classified as seronegative if it has titre < 40 to all 4 DENV serotpyes; monotypic if it has titre ≥ 40 to one DENV serotype only or if it has titre ≥ 40 to more than one DENV serotypes with ≥ 6 fold higher response of the dominant (i.e. highest) DENV serotype versus the second dominant DENV serotype; multitypic if it has titre ≥ 40 to more than one DENV serotypes with < 6 fold higher response of the dominant (i.e. highest) DENV serotype versus the second dominant DENV serotype. For monotypic subjects, the dominant DENV serotype is assumed to be the serotype of the previous DENV infection. Figure S20 shows the mean and 95% confidence interval (CI) of the observed DENV titres and raises in titres in time in DENV seronegative, monotypic and multitypic profiles, according to the definitions given in this section. Figures S21 and S22 show the mean and 95% confidence interval (CI) of the observed DENV titres in time and of the observed rises in titres between successive vaccinations respectively, having classified the subjects as DENV seronegative, monotypic profile with detected DENV1 titre, monotypic profile with detected DENV2 titre, monotypic profile with detected DENV3 titre, monotypic profile with detected DENV4 titre or multitypic infection profile, according to the definitions given in this section. , post-dose 2 (PD2) and post-dose 3 (PD3) titres (row 1) and rises in titres (row 2) for each DENV serotype (columns) by baseline immunological status against DENV (colour code). Here, DENV-denotes subjects with titre <40 for all DENV serotypes; monotypic denotes subjects with titre ≥ 40 to only one DENV serotype or with titre ≥ 40 to more than one DENV serotype and ≥ 6 higher response of the dominant titre versus the second highest titre; multitypic denotes subjects with titre ≥ 40 for at least two DENV serotypes and < 6 higher response of the dominant titre versus the second highest titre. Undetectable titres (i.e. titres <10) are assigned a titre value of 5. Titres are shown on a log10 scale. Increases in titres are shown on a log2 scale, according to the definition.

Comparison between definitions
Tables S19 and S20 summarise the baseline immunological characteristics of the subjects included in the descriptive analysis according to the definitions given in sections 3.3.1-3.3.4. The baseline immunological characteristics of the subjects included in the descriptive analysis according to definition 1 are given in Table   1 of the main text.   Table S20: Baseline immunological characteristics of the subjects included in the descriptive analysis according to definition 3 with threshold at 10 and 40 given in sections 3.3.3 and 3.3.4, respectively.