Effect of biannual azithromycin distribution on antibody responses to malaria, bacterial, and protozoan pathogens in Niger

The MORDOR trial in Niger, Malawi, and Tanzania found that biannual mass distribution of azithromycin to children younger than 5 years led to a 13.5% reduction in all-cause mortality (NCT02048007). To help elucidate the mechanism for mortality reduction, we report IgG responses to 11 malaria, bacterial, and protozoan pathogens using a multiplex bead assay in pre-specified substudy of 30 communities in the rural Niger placebo-controlled trial over a three-year period (n = 5642 blood specimens, n = 3814 children ages 1–59 months). Mass azithromycin reduces Campylobacter spp. force of infection by 29% (hazard ratio = 0.71, 95% CI: 0.56, 0.89; P = 0.004) but serological measures show no significant differences between groups for other pathogens against a backdrop of high transmission. Results align with a recent microbiome study in the communities. Given significant sequelae of Campylobacter infection among preschool aged children, our results support an important mechanism through which biannual mass distribution of azithromycin likely reduces mortality in Niger. The trial protocol was reviewed and approved by the Committee on Human Research at the University of California, San Francisco, and the Niger Ministry of Health’s Ethical Committee. Parents or guardians of enrolled children provided oral consent before each azithromycin or placebo treatment and at each specimen collection visit. Parents or guardians were instructed to report any adverse event within 7 days of treatment by contacting their village representative, who then reported events to the site coordinator and UCSF. An independent Data and Safety Monitoring Committee provided additional oversight. Centers for Disease Control and Prevention (CDC) researchers had access to de-identified samples for analysis (no personally identifying information). MORDOR Niger was a cluster-randomized, placebo-controlled trial that randomized at the community level because of the intervention’s campaign-style, biannual mass distribution. Communities with 200–2000 inhabitants based on the Niger 2012 census were eligible for inclusion in the trial, and children ages 1–59 months who weighed >3.8 kg were eligible for treatment. An intensive morbidity monitoring trial enrolled 30 communities and randomized them 1:1 to receive either biannual azithromycin or placebo to all children 1–59 months old ({“type”:”clinical-trial”,”attrs”:{“text”:”NCT02048007″,”term_id”:”NCT02048007″}}NCT02048007). The trial used a repeated cross-sectional design, whereby 40 children per community were randomly sampled in each measurement round and invited to participate in a monitoring visit. The trial’s open cohort design meant that children aged in and out of the study based on their age at the time of treatment. Field staff collected dried blood spots from participating children at baseline and annually thereafter at 12, 24, and 36 months of follow-up. The antibody substudy included a supplemental visit at 6 months, following the malaria season. Children who were randomly selected in multiple survey rounds contributed to longitudinal analyses. Field data were collected using handheld tablets (Android operating system version 5) and a custom application designed for the study (versions 2–4, Conexus Inc., Los Gatos, CA), which encrypted and transmitted the data to a secure server hosted by Salesforce.com. Communities were randomized 1:1 using a sequence the trial biostatistician (TCP) generated. Unmasked members of the data team and Pfizer labeled the study drugs. Placebo and azithromycin had identical packaging to maintain masking. Participants, field staff, laboratory staff, analysts, and all investigators were masked to treatment assignments throughout the trial. Masked analyses were completed using a shuffled version of the treatment assignment variable39. Data were unmasked only after the final table and figure shells had been populated (documented through the article’s GitHub repository). Children ages 1–59 months received azithromycin or identically-appearing placebo at the time of enrollment and every 6 months over the course of the study through community-wide census and MDA distributions performed by study staff. Children were given a volume of oral suspension equal to at least 20 mg per kilogram of body weight, which was measured by hanging scale for children unable to stand or by height stick for children who could stand, consistent with the Niger trachoma program. Children with a known allergy to macrolides did not receive azithromycin or a placebo. Dried fingerprick blood spots (DBS) were collected onto calibrated filter paper wheels with six 10 µl extensions (TropBio Pty Ltd., Townsville, Queensland, Australia), which were dried and packaged in individual sealable plastic bags with desiccant and stored in a −20 °C freezer prior to shipping to CDC. DBS were shipped to CDC at ambient temperature and stored in a −20 °C freezer40. A single 10 µl extension per participant was eluted overnight at 4 °C in phosphate-buffered saline (PBS) containing 0.5% casein, 0.3% Tween-20, 0.5% polyvinyl alcohol, 0.8% polyvinylpyrrolidone, 0.02% sodium azide, and 3 µg/mL E. coli extract (Buffer B). Elutes were diluted to a final concentration of 1:400 with additional Buffer B to test on the multiplex bead assay (below). Antigens were covalently coupled to polystyrene beads (SeroMap Beads; Luminex Corporation, Austin, TX) by modifying carboxyl groups on the beads to ester groups using 1-ethyl-3-(3-dimethylaminopropyl) carbodiimide (EDC) (EMD Millipore Calbiochem, USA) in the presence of N-hydroxysulfosuccinimide (sulfo-NHS) (Thermo Scientific Pierce, USA)40. Ester groups on the beads bind to primary amine groups on antigens to create stable amide covalent bonds. Beads were briefly sonicated in a water bath and washed with 0.1 M sodium phosphate buffer, pH 6.2 (NaP) in preparation for bead activation. Beads were protected from light and rotated for 20 min in NaP with 5 mg/ml each EDC and NHS. After activation, activated beads were washed and suspended in coupling buffer, antigen added, and rotated at room temperature for 2 h. Coupling buffers and antigen amounts were previously determined for each antigen (Supplementary Table 5). After 2 h, antigen-coupled beads were washed with PBS and unreacted sites were blocked with 1% bovine serum albumin (BSA) in PBS for 30 min. Antigen-coupled beads were resuspended in storage buffer (PBS, 1% BSA, 0.05% Tween-20, 0.02% NaN3, protease inhibitors) and kept at 4 °C until use in assays. The panel included malaria antigens to various stages of P. falciparum infection, including sporozoite (CSP), hepatocyte (LSA1), merozoite (GLURP-R0) and erythrocyte (MSP-119, AMA1, HRP2), with MSP-119 and AMA1 thought to induce longer-lived IgG responses compared with other included P. falciparum antigens12. Species specific MSP-119 was used to detect P. vivax, P. malariae, P. ovale10. Bacterial and protozoan antigens from Campylobacter jejuni (p18, p39), enterotoxigenic Escherichia coli labile toxin B subunit (ETEC LTB), Vibrio cholerae toxin B subunit (CTB), Salmonella serogroups B and D (LPS), Cryptosporidium parvum (Cp17, Cp23), Giardia duodenalis (VSP-3, VSP-5), and Streptococcus pyogenes serogroup A Pyrogenic Exotoxin B (SPEB) were also coupled to beads9,41. We measured IgG responses using a multiplex bead assay on the Luminex platform. Antigen-coupled beads (1250 per well/bead coupling) were incubated in 96-well assay plates with diluted sample for 1.5 h then washed with 0.3% Tween-20 in PBS (PBST). Beads incubated with biotinylated mouse anti-human IgG (1:500 dilution) and biotinylated mouse anti-human IgG4 (1:1250 dilution) for 45 min to detect IgG bound to the beads. After additional washes with PBST, beads were incubated for 30 min with phycoerythrin-labeled streptavidin (1:200 dilution) to detect bound biotinylated anti-human IgG. After detection, beads were washed with PBST and incubated for 30 min with PBS containing 0.5% BSA, 0.05% Tween-20, and 0.02% sodium azide to remove loosely bound antibodies. After a final wash with PBST, beads were resuspended in PBS and stored at 4 °C overnight. The next day, assay plates were read on a Bio-Plex 200 instrument (Bio-Rad, Hercules, CA) equipped with Bio-Plex manager 6.0 software (Bio-Rad). The median fluorescence intensity (MFI) with the background from the blank well (Buffer B alone) subtracted out (MFI-bg) was recorded for each antigen for each sample. DBS samples were masked and randomly ordered by the UCSF trial coordinating center before sending them to the laboratory at the CDC. Samples were run on 63 plates and each plate included positive controls from a high positive sera pool (1:400 dilution), a low positive sera pool (1:6400 dilution), and a normal human sera (single) control. To assess plate-to-plate variation, we estimated the standard deviation (SD) for each antigen’s responses across plates. The laboratory protocol specified that a plate should be re-analyzed if more than half of the antigens had SDs that deviated by >20% from the overall average in multiple controls. Two of 63 plates failed this criterion for one control but passed in the other two controls and were thus not repeated. For the positive control sample responses to the 22 antigens used in this study, the average CV% was 10.0 with an SD of 4.9. The median CV% was 9.3 with a range of 2.9–20.2%. At the time of dried blood spot collection, a drop of blood was used to prepare a thick smear slide that was stained with 3% Giemsa and later assessed for the presence and density of malaria parasites4. Lab personnel were masked to specimen groups, and two independent team members at the Centre de Recherche Médicale et Sanitaire (Niamey, Niger) with experience in parasitemia assessment recorded the presence and density of parasites. Any discrepancies were adjudicated by a third, masked, independent laboratory worker until they reached a majority consensus. We estimated community-level malaria parasitemia as the proportion of children with positive thick smears by microscopy. We compared groups using geometric mean IgG responses, seroprevalence, and the seroconversion rate, including measurements at all follow-up times (6, 12, 24, and 36 months). These were pre-specified, secondary outcomes for the trial ({“type”:”clinical-trial”,”attrs”:{“text”:”NCT02048007″,”term_id”:”NCT02048007″}}NCT02048007). We log10 transformed Luminex MFI-bg IgG levels before analysis. We converted IgG responses to seropositive and seronegative classes using seropositivity cutoffs derived from the mean plus 3 SD of responses from a panel of 92 anonymous, adult, USA resident blood donors (all malaria antigens), from ROC-derived cutoffs based on responses from known positive and negative specimens from North America (Cryptosporidium n = 68, Giardia n = 32)42, or from the mean plus 3 SD of presumed unexposed measurements (all other antigens). We identified presumed unexposed measurements as those collected among children ≤12 months old that preceded a 10-fold increase in IgG in the longitudinal subsample9. For pathogens with multiple antigens measured, we classified children as seropositive if they were positive to any antigen. For P. falciparum, we examined individual antibody endpoints as well as a composite outcome, defined as a seropositive response to any P. falciparum antigen measured. An exploratory analysis (not pre-specified) grouped P. falciparum responses by antigens with more durable IgG responses (MSP-1, AMA1) and less durable IgG responses (GLRUP-Ro, LSA1, CSP, HRP2). We restricted the age ranges included in the analyses based on pre-specified rules to exclude maternal IgG contributions and to focus the analysis on age ranges with heterogeneity in IgG responses. Before data were unmasked, we examined age-antibody profiles for each antigen and excluded from the primary analyses measurements among children <12 months (malaria responses) and <6 months (bacterial and protozoan responses) to remove potential maternally derived IgG contributions (SI Fig. 3, SI Fig. 8)43. Additionally, we limited all analyses of ETEC LTB to ages 6–24 months and force of infection analyses based on seroconversion rates to measurements among children ≤24 months (all enterics except Salmonella sp.) because nearly all children older than 24 months were seropositive. We summarized the SD of community-level seroprevalence and estimated the ICC for community-level responses using a mixed-effects binomial model with a parametric bootstrap to estimate 95% confidence intervals for the ICC (1000 iterations)44. We compared community-level malaria parasitemia prevalence and seroprevalence to P. falciparum antigens using Spearman rank correlation. All comparisons were intention-to-treat. We compared mean differences between groups in geometric mean IgG levels, seroprevalence, and malaria parasitemia by pooling all post-treatment measurements. We estimated 95% confidence intervals using a non-parametric bootstrap that resampled communities with replacement (1000 iterations). We calculated exact permutation P-values from the randomization distribution of mean differences. We used a current status, semi-parametric proportional hazards model to estimate the force of infection from age-structured seroprevalence45. We fit a generalized additive mixed model with binomial errors and complementary log–log link where Yij is antibody seropositivity, Aij is the age for child j in community i. Xi is treatment allocation for community i (equal to 1 for azithromycin, 0 for placebo). The model included community-level random effects, bi, to allow for correlated outcomes within the community. Function g(·) was parameterized with cubic splines that had smoothing parameters chosen through generalized cross-validation using the default in the R mgcv package45. The primary analysis pooled information over all post-randomization measurements available at the time of analysis (months 6, 12, 24 and 36). We estimated the hazard ratio (HR) of seroconversion associated with the biannual mass distribution of azithromycin as θ^HR=exp(β^1). We estimated age- and treatment-specific seroprevalence from the model as and we estimated age- and treatment-specific force of infection from the model as where η′(a, x) is the first derivative of the linear predictor from the complementary log–log model, η(a, x)46. We estimated η′–(a, x) using finite differences from the model predictions45,47. We estimated approximate, simultaneous 95% confidence intervals around age-specific seroprevalence and age-specific force of infection curves with a parametric bootstrap (10,000 replicates) from posterior estimates of the model parameter covariance matrix48. We used the age-specific force of infection curves to visually confirm proportional hazards between groups. To estimate the marginal average force of infection in each group, we integrated overage49. In some cases, the same child was measured multiple times during the trial and contributed multiple measures within a community over the course of follow-up (next section). We considered the inclusion of a child-level random effect in the age-structured seroprevalence model, nested within the community, to model this additional potential source of outcome correlation. However, models that included this more complex random effects structure failed to converge in most cases. Since our target of inference was the HR for azithromycin-treated communities versus placebo, inclusion of a community-level random effect, the independent unit of analysis and inference, should lead to unbiased estimates of the standard error even without attempting to model additional sources of variation within the community50. A subset of children was sampled in multiple, repeated cross-sectional surveys and thus provided longitudinal antibody measurements (two to five visits). Children included in multiple cross-sectional samples between ages 12 and 59 months contributed to longitudinal analyses of malaria (919 children, 2197 measurements, median [range] measurements per child 2 [2, 5]). Longitudinal samples from children ages 6 to 59 months (1038 children, 2516 measurements, median [range] measurements per child 2 [2, 5]) contributed to analyses of Salmonella and Streptococcus, and longitudinal samples among children 6–24 months (313 children, 680 measurements, median [range] measurements per child 2 [2, 4]) contributed to analyses of the remaining enteric pathogens. We conducted a supplemental analysis in this opportunistic subgroup to estimate prospective seroconversion and seroreversion rates. We defined seroconversion as an increase in IgG MFI-bg to a level above the antibody’s seropositivity cutoff. For pathogens with multiple measured antigens, a child was deemed to have seroconverted if either antibody response met the definition for seroconversion. We assumed that seroconversions occurred at the midpoint of the interval between measurements when estimating person-time at risk. We jointly estimated seroreversion rates using the same approach, using a decrease in IgG across the seropositivity cutoff. We used a non-parametric bootstrap, resampling clusters with replacement, to estimate 95% confidence intervals for rate and incidence rate ratio estimates. We pre-specified examining treatment differences by age at the trial’s start date (<6 months vs. older) and by trial phase (6, 12, 24, and 36). We hypothesized that azithromycin could reduce antibody-based measures of transmission more among younger children who were immunologically naïve, and at later phases of the trial due to additional rounds of biannual MDA. For the age-stratified subgroup, we focused on children 6 months or younger (including not yet born) at the start of the trial because age 6 months is approximately when we would expect children to begin to produce their own IgG response (additional details in the statistical analysis plan)43. We omitted a pre-specified subgroup analysis for the rainy versus dry season because we ultimately determined that too few samples were collected after the rainy season (6-month samples only). An exploratory subgroup analysis stratified of P. falciparum IgG comparisons by baseline malaria parasitemia (not pre-specified). We estimated the proportion of children who were malaria positive by thick smear microscopy in the baseline survey and identified a natural break in the distribution with 17 communities ≤5% parasitemia prevalence and 13 communities >5% parasitemia prevalence. We assessed the sensitivity of the results to the seropositivity cutoff for each antigen by changing cutoff values ±20% and then repeating the primary analyses. We assessed the sensitivity of the results to individual communities by conducting a leave-one-out analysis that examined the distribution of analysis results excluding each community in turn and computed a jackknife bias estimate for the difference in seroprevalence and the log HR of seroconversion rates51. The MORDOR morbidity monitoring trial was designed around the primary antimicrobial resistance monitoring endpoints5,52. For the present analyses, assuming a sample of 15 communities per arm and 140 measurements per community over four rounds, the seroprevalence of 65% (P. falciparum MSP-1), a community-level ICC of 0.004, and a two-sided alpha of 5%, we estimated that we would have 80% power to detect a reduction of 5.4 percentage points in seroprevalence due to intervention53 (the pre-analysis plan provides additional details, https://osf.io/d9s4t/, seroprevalence, and ICC assumptions were estimated from the nearby PRET trial)54. Within each set of analyses, we estimated P-values adjusted for multiple comparisons allowing for a 5% false-discovery rate using the Benjamini–Hochberg correction55. Further information on research design is available in the Nature Research Reporting Summary linked to this article.