In order to conduct a sensitive, high-throughput screen for superinfection in the Mombasa cohort, we developed a pipeline for amplification, next-generation sequencing (NGS), data cleaning, and phylogenetic and sequence diversity analysis of longitudinal plasma RNA (Fig. 1). One-hundred thirty-two women met our selection criteria for the NGS superinfection screen, with a median follow-up time of 2046 days (IQR 1265–2848). We successfully amplified gag, pol and env at three timepoints in 115 women and at least two genomic regions in at least the first and last timepoints in 129. The remaining 3 women were dropped from analysis. In total, ∼1.7 million raw sequencing reads were obtained, with ∼1.25 million passing quality filtering: a median of 901 per amplicon per sample.

Women were considered putative superinfection cases if the posterior probability of monophyly supported single infection at the earliest studied timepoint followed by introduction of a distinct viral clade and increased viral diversity consistent with that seen in simulated dual infection (Fig. 1e&f). Putative cases of superinfection were confirmed and their timing specified by analyzing intervening timepoints. Nine cases of superinfection were detected and their timing specified. One case of suspected dual infection was detected, in which two clades were detected at the earliest sample analyzed (60 days post-infection (dpi)) and throughout infection (data not shown).

Example data from two cases of superinfection are summarized in Figures 2 and 3. Initial screening of subject QD151 (Fig. 2) showed monophyletic subtype A infection at 39 dpi and two subtype A clades in all three genes at 938 and 1701 dpi. In subsequent analysis of intervening timepoints the second clade was first detectable at 801 dpi (Fig. 2a). At this time, pairwise distance increased sharply, for example in gag from 0.27% at 241 dpi, to 12.75% at 801 dpi (Fig. 2b), into the range observed in simulated dual infections. These observations supported introduction of a second subtype A variant between 241 and 801 dpi. The initial clade was no longer detectable in pol at 1701 dpi, suggestive of a genomic recombination event (Fig. 2c). Similarly, subject QB210 (Fig. 3) showed initially monophyletic infection with a subtype A/D virus, followed by introduction of a subtype C/D virus at 163 dpi, evidenced by polyphyly and a shift in pairwise distance (>10%) in all 3 genes (Fig. 3a and 3b). In intervening timepoints, the second variant could be detected in all genes at 163–170 dpi, but was undetectable in gag and pol after 170 dpi, indicating recombination (Fig. 3c).

Characteristics of the 9 new cases of superinfection are summarized in Table 1 and Figure S2. In all but two cases the superinfecting variant was detected in all 3 amplicons in at least one timepoint. In all cases, the superinfecting variant was detected at multiple timepoints in at least one amplicon. In one case (QC369), the initial variant became undetectable in any amplicon following superinfection, suggesting it was replaced, to our detection limit, by the superinfecting variant. Both variants were detected at two timepoints each, the initial variant at 17 dpi and 28 dpi, and the superinfecting variant at 143 dpi and 451 dpi (Fig. S2), indicating this result was not due to contamination. Further, the possibility of sample mix-up was excluded by HLA-typing (data not shown). As illustrated in Figures 2, 3 and S2, in the other 8 cases, variants were intermittently detected in different amplicons at different times, suggestive of genomic recombination and dynamic turnover of the circulating viral population.

Combining the data here with those from previous studies in the Mombasa cohort [5], [12], [17], a total of 146 women were examined for superinfection: 90 were tested using NGS, 39 using both NGS and Sanger sequencing, and 17 using only Sanger sequencing. Among the 39 women previously identified as singly infected by Sanger sequencing and tested by NGS here, no new cases of superinfection were identified, suggesting older methods were sensitive enough to detect superinfection. Twenty-one cases of superinfection were confirmed based on detection of the superinfecting virus in two or more samples. The timing windows of all 21 superinfection events are summarized in Figure 4 and Table S2. The midpoint of the timing window of the 9 new cases ranged from 81 to 1041 dpi, with 6 occurring within the first year of infection. The window of superinfection events was defined to a median of within 127 days, with window sizes of 90 to 1253 days. Timing of all 21 cases ranged from 63 to 1895 dpi, defined to a median of within 146 days.

We detected both inter-subtype and intra-subtype superinfections. In 6 of 9 cases identified by NGS, the superinfecting variant was the same subtype as the initial variant in every gene where both were detected. In all 9 cases, the variants were the same subtype in the env amplicon (Table 1). Among all 21 cases of superinfection (Table S2), the majority of superinfection events we detected were intrasubtype, regardless of genomic region: 53.8% were intrasubtype based on gag sequence, 62.5% based on pol, and 70.6% based on env.

We further investigated the possibility of a bias in sequence similarity of superinfecting variants to initial variants by analyzing amino acid diversity. We compared the pairwise amino acid distance between initial and superinfecting variants within each superinfection case to the distance that would be expected by chance. The latter was modeled by simulated mixtures of sequences from all possible pairs of singly infected individuals in the Mombasa cohort (Fig. 5). Using NGS data from the 9 superinfection cases and 120 singly infected women screened here, we found no significant differences between the sequence similarity within superinfected individuals and that expected by chance (Fig. 5a). Including Sanger sequencing data from the additional 12 superinfected women previously screened yielded a similar result (Fig. 5b)

The incidence of superinfection among women who were screened was compared to the incidence of initial infection in the entire cohort at risk. Only incident HIV infections (occurring after enrollment in the cohort) were included. Fourteen women who were seronegative but HIV RNA positive at enrollment were excluded for this reason. Seven of these had been screened for superinfection, and one was found to be superinfected, which mirrors the frequency of superinfection observed in the entire group. The individual with evidence of dual infection at the earliest timepoint was also excluded, since we were unable to distinguish coinfection from superinfection. After exclusions, 1910 women were at risk of initial infection, contributing 5124 person-years, and 138 women were screened for superinfection, contributing 764py following first infection. There were 295 initial infections, giving a crude incidence rate of 5.7 per 100pys, and 20 superinfections, giving a crude incidence rate of 2.61 per 100 pys.

The incidence of superinfection and initial infection over time is summarized in Figure 6. We used Andersen-Gill proportional hazards analysis to generate a hazard ratio (HR) relating the incidence of superinfection to that of initial infection. The unadjusted HR for this comparison was 0.49 (CI 0.31–0.76, p = 0.0018). Variables previously shown to influence HIV exposure risk in this cohort [22], [23] were included as adjustments in the model (summarized in Table 2). These included self-reported sexual risk behavior, place of work, hormonal contraceptive use, genital tract infections, years in sexwork, age at first sex, total follow-up time in the cohort and calendar year. The HR for superinfection compared to initial infection, adjusted for these variables, was 0.47 (CI 0.29–0.75, p = 0.0019). Since proportional hazards analysis is based on time to infection and the precision with which superinfection timing was determined varied between cases, we performed sensitivity analyses setting infection timing for all cases to the start or midpoint of the timing windows rather than the end, as done for the above analysis. In both of these analyses, significant differences in incidence were also observed: setting infection timing to the start of the windows, the adjusted HR was 0.33 (CI 0.18–0.58, p = 0.00012); using the window midpoints, the adjusted HR was 0.39 (CI 0.23–0.63, p = 0.00016).

We assessed whether the risk of superinfection varied with time since initial infection by dividing our data into infection events occurring early or late in follow-up and estimating the HR, as above, in each subset. We found that within the first 6 months at risk, the incidence rates of initial and superinfection did not differ significantly (adjusted HR 0.73, p = 0.51), whereas after 6 months the rate of superinfection was lower than that of initial infection (adjusted HR 0.40, p = 0.0017). A similar result was observed when considering events within or beyond one year at risk: within the first year, the incidence rates of initial and superinfection did not differ significantly (adjusted HR 0.54, p = 0.14), but after one year the rate of superinfection was significantly lower (adjusted HR 0.43, p = 0.0059). Sensitivity analyses setting infection time to the start and midpoint of the timing windows as above reproduced the same results (data not shown).

We noted that the previous screens in the cohort appeared to detect a higher frequency of superinfection than the NGS screen (12 cases of 56 women screened, compared with 9 cases of 90), with a greater fraction of the events occurring later after initial infection (Fig. 4). Since the NGS screen spanned later years in the cohort than the previous studies, such a difference could be due to the known decline in infection risk in the cohort over calendar time [22], [23]. However, the numbers of events are small when the datasets are considered separately and the difference both in superinfection incidence rate and post-infection timing between the two studies was not statistically significant (data not shown).

The study was approved by the ethical review committees of the University of Nairobi, the University of Washington and the Fred Hutchinson Cancer Research Center. Written informed consent was obtained from all participants.

Seronegative women in Mombasa, Kenya, attended monthly visits, at which clinical examinations, interviews and sample collection took place, as previously described [22]. Following seroconversion, sample collection took place quarterly. Individuals were selected for superinfection screening based on sample availability <6 months and >2 years post-initial HIV infection, and an approximately equally spaced intervening sample. Within these limitations, samples with maximal plasma viral load, >1000 copies/ml, and prior to initiation of antiretroviral therapy were selected. Thirty-nine of 44 women previously screened for superinfection by Sanger sequencing and identified as singly infected [5], [12], [17] were rescreened; the remaining 5 women did not have adequate samples available.

HIV virions were isolated from heparinized plasma using the μMACS VitalVirus HIV Isolation kit (Miltenyi Biotec) and viral RNA extracted from 140–420 µl, depending on viral load, using the Qiamp viral RNA Mini kit (Qiagen). Nested RT-PCR of ∼500 bp in gag, pol and env was conducted in duplicate (see Table S1). RNA input into each reaction was normalized to 3000 viral genomes according to plasma viral load, or the maximum possible where viral load was too low. RT-PCRs for the three genes were multiplexed. Nested PCR reactions were carried out separately for each region with primers containing adaptors for Roche 454 sequencing and a unique 8 bp barcode sequence to identify each sample. PCR products were purified using AMPure XP PCR purification beads (Agencourt) and quantified using the Qubit dsDNA HS assay (Invitrogen). PCR products were sequenced on the Roche 454 GS-Junior or GS-FLX titanium platform. Where initial sequencing suggested superinfection (see below), timing was inferred by sequencing intervening timepoints. Sequences are available upon request from the authors.

454 sequences were error-corrected using AmpliconNoise [32]. Chimeric sequences were identified and removed using UCHIME [33]. Cross-contamination between samples sequenced together and contamination by other lab samples was identified by all-against-all BLAST against a local database of published HIV sequences and sequences from the same sequencing run. Sequences with high identity hits to known laboratory stains or other samples from the same sequencing run were removed. Sequences with abundance <5 reads or 0.5% of the sample, whichever was higher, were excluded from further analyses as lower abundance variants were not reproducibly detected in repeated deeper sequencing of two selected samples where rare variants formed a distinct phylogenetic clade. An amplicon-specific profile HMM was created from an alignment of representative sequences from multiple subtypes. For each subject and amplicon, 20 reference sequences were selected by placing 454 reads on a tree of candidate reference sequences [34] and minimizing the average distance to the closest leaf [35]. These reference sequences, representatives from subtypes common to the region, and 454 reads were aligned to the HMM using hmmalign [36] and non-consensus columns removed. Any sequences <200 bp long after alignment and trimming were removed. We used BEAST [37] to calculate a posterior probability of monophyly for the sequences. A posterior sample of trees was obtained using a strict molecular clock, Bayesian Skyline Plot population model and the HKY substitution model. Each MCMC chain ran 20 million iterations, sampling every 2000, discarding the initial 25% of samples as burn-in. Chains were assessed for convergence by examining effective sample size (ESS) and by visual inspection of traces of key parameters. A strict clock was used as poor mixing was frequently observed under relaxed clock models. BEAST runs with intermediate posterior probabilities (0.2–0.8) were manually examined for recombinant sequences and run again with putative recombinants removed. Pairwise distances were calculated for all sequence pairs under the TN93 model using APE [38], reporting the maximum within-subject distance. For comparison, 95% confidence limits of pairwise distances were calculated for sequences from known single infections (previously screened in [5], [12], [17]) and simulated dual infections. Dual infections were simulated by combining all pairs of sequences from previously screened singly infected samples. Pairwise distances calculated from 454 sequences obtained in this study were compared to the upper bound of the 95% quantile of single infection distances, and the lower bound of the 95% quantile of simulated dual infection distances. This pipeline was validated and refined by processing monophyletic viral isolates, known mixtures of isolates, and known cases of superinfection detected by Sanger sequencing [17]. These methods were found to be sensitive enough to distinguish two subtype A isolates mixed at abundances of 5%∶95% genome copies in all three genomic regions, and at 1%∶99% in two of three genomic regions (Fig. S1).

Sequences were aligned as for the phylogenetic analysis. Insertions relative to the reference alignment were removed, and sequences with <60% coverage or identified as recombinants between initial and superinfecting variants upon visual inspection were excluded. For each case of superinfection, viral sequences were annotated as the initial strain or the superinfecting strain. We calculated the mean Hamming distance between amino acid sequences of the superinfecting strain from the time of superinfection detection and sequences of the initial strain up to and including this time. In calculating the mean distance, each pairwise comparison was weighted using the product of the multiplicities of the two reads. To investigate whether these distances deviated from what would be expected by chance, an artificial set of mock superinfections was generated by combining sequences from singly infected individuals. All pairs of singly infected individuals screened by 454 sequencing were enumerated. In each pair, one individual was randomly chosen to be the source of the ‘initial’ virus in the simulated superinfection. A time of ‘superinfection’ was chosen randomly from the available sampled timepoints and sequences from all timepoints up to and including this time were used for analysis. The other individual in the pair acted as the source of the ‘superinfecting’ virus. A time of ‘transmission’ was chosen randomly from the available sampled timepoints and sequences from this timepoint were used. Mean distances within pairs were calculated as above. The analysis was repeated including gag and env Sanger sequences from previously published cases [5], [12], [17], trimmed to the genome region amplified for NGS, and given unit weight. A two-sample Wilcoxon test was used to test for a difference between the distances observed in true superinfections and those simulated in mock superinfections.

Statistical analysis was performed using R (www.r-project.org). The incidences of initial and superinfection were compared by Andersen-Gill proportional hazards analysis. The predictor was inclusion in the screen for superinfection, modeled as a time-dependent variable, and the outcome was time to HIV infection (initial and super). Timing of infection events for the incidence analysis was set to the study visit of their detection (for initial infection events the visit after inferred infection timing; for superinfection events, the time at which the superinfecting virus was first detected). Individuals who were HIV infected but not screened for superinfection were censored after acquisition of initial infection. Individuals who became superinfected were censored after acquisition of superinfection. Individuals who were screened and not found to be superinfected were censored at the last timepoint screened. Since samples after initiation of antiretroviral treatment were excluded from superinfection screening, no follow-up after treatment initiation was included. The model was adjusted for time-varying variables at each visit: calendar year, age, years in sexwork, number of weekly sexual partners, number of weekly unprotected sex acts, hormonal contraceptive use in the prior 70 days and any genital tract infection in the prior 70 days (bacterial vaginosis, cervicitis, genital ulcer disease, gonorrhea, trichomoniasis); place of work and age at first sex recorded at enrollment; and total follow-up time in the study. Incidences of initial and superinfection were also estimated as described in [8], using Poisson regression and propensity score matching to select a subset of women at risk of initial infection whose baseline risk profiles most closely matched those of women screened for superinfection.