Characterizing the Temporal Dynamics of Human Papillomavirus DNA Detectability Using Short-Interval Sampling

Outcome assessment in human papillomavirus (HPV) natural history studies relies on viral DNA detection from cervical samples collected at fixed intervals, usually every 4 to 6 months. Several recent studies using highly sensitive PCR-based tests to measure HPV DNA have reported that recurrent detection of an HPV type after a period of nondetection is common, ranging from around 8% to 19.4% (1–3). Weaver and colleagues reported that samples testing HPV16-negative in between runs of HPV16 DNA positivity by standard consensus primer PCR (linear array) were found to be HPV16-positive when more sensitive type-specific amplification assays were used, suggesting that recurrent detection is more likely to represent viral load fluctuations of a persistent infection or periods of latency and reactivation, rather than reinfection of a previously cleared viral type (4).

Because HPV incidence rates are based on person-time at risk for new infection, accurate estimates require valid assessment of baseline infection status at enrollment into a natural history study. Because HPV natural history studies suffer from left truncation bias (unobserved HPV DNA status before study entry) and interval sampling (unobserved HPV DNA status between study visits or sample collections), it is likely that a substantial fraction of newly detected HPV in observational studies is misclassified as an initial infection, rather than recurrent detection of a persistent infection (5).

In this study, we sought to characterize the temporal variability of detectable HPV DNA based on frequent self-sampling by women over a 16-week period. Under an assumption that fluctuation in DNA detection was largely a result of variable detection in women with low viral load (4), we estimated biases between prevalence estimates based on a single versus multiple samples using the 16-week cumulative prevalence as the “true” prevalence. Under an alternative assumption that the loss and gain of DNA detection reflected acquisition, clearance, and reinfection or deposition of viral DNA from a male partner, we investigated whether recurrent detection or its subsequent loss of detection was associated with sexual activity at follow-up. In the absence of molecular markers to assess the validity of each hypothesis tested, these analyses were intended to evaluate the alternative assumptions for the best fit to the observed data.

We used archived vaginal swabs from a prior study of vaginal douching cessation. The study design and characteristics of the study population were previously described elsewhere (6). In brief, the 16-week study began with a 4-week observational period (phase I) followed by a 12-week douching cessation where women were asked to withhold their douching practice (phase II). Forty-seven eligible women were enrolled between December 2005 and March 2007. Ethical approval for the primary study was obtained from the Institutional Review Boards (IRB) of the Johns Hopkins University School of Medicine (Baltimore, MD). There was no evidence suggesting any effect of douching on HPV DNA detectability (7), and we did not observe any difference in HPV detection between the two study phases within the same subjects. Therefore, the current analysis included women who completed both phases I and II of the original study.

After consenting, women were interviewed at baseline and at the end of phase II to collect demographic information such as age, race, highest education levels, and reproductive health history including oral contraceptive use, sexual partnerships, and behaviors. Women were also asked to keep daily diaries on menstrual bleeding and sexual activity during the 16-week follow-up. In both phases, participants self-collected mid-vaginal samples twice per week (every 3–4 days). Instructions were provided at baseline on how to perform the self-sampling and to transport the swabs back to the study laboratory. DNA was extracted from the vaginal swab, as previously described (8), and genotyped using the Roche HPV Linear Array Kit according to the manufacturer's instructions. The detection limit of linear array is 10 to 100 viral copies for each of the genotypes tested. Each PCR test included 81 study specimens, four characterized positive controls (25 and 100 copies of HPV 16 and 18 plasmid DNA in a background of roughly 30,000 human cellular DNA equivalents); the linear array kit extraction controls: HPV DNA–negative (containing 1.0 × 10⁶ K562 cells/mL), high HPV DNA–positive (1.25 × 10⁵ SiHa cells and 1.0 × 10⁶ K562 cells/mL), and low HPV DNA–positive controls (5.0 × 10³ SiHa cells and 1.0 × 10⁶ K562 cells/mL); six negative (no DNA) controls (9). A PCR batch was considered valid if all positive controls were positive for human β-globin and the intended PCR genotype (and no others) and the negative controls were completely negative. Any deviation would result in an invalid run and the batch would be repeated. All runs included in the current study were valid.

Comparisons of demographic and behavioral characteristics of the study population were achieved by Wilcoxon rank-sum test for continuous variables and by χ² test or Fisher exact test for categorical variables. HPV DNA positivity at the woman-level (for any type) and at the genotype-level (for a specific type) was evaluated at each prespecified sampling time (point prevalence) and over the study period (cumulative prevalence), respectively. Numbers of detected viral types at each sampling time were summarized graphically for each individual over time (10). In addition, type-specific temporal variability in viral detection was visualized using history transition matrices (11).

Assuming that the cumulative prevalence reflected the underlying infection status, absolute bias was calculated by taking the difference between the baseline and the cumulative prevalence, whereas relative bias was determined by the ratio of baseline to cumulative prevalence minus 100%. The baseline result was excluded from the cumulative prevalence while performing the exact McNemar test to examine whether point and cumulative prevalence estimates were statistically different from each other. Applying the bootstrapping technique, the uncertainty surrounding, as well as the representativeness of, the baseline estimate was further quantified by randomly assigning a single time point during the 16-week period as the baseline and then obtaining the corresponding bias estimates. This procedure was repeated 400 times to determine the distribution of bias estimates for prevalence of any type and of high-risk types. In addition, 100,000 bootstraps were performed to determine the minimal number of consecutive swabs needed from the same woman to reduce the median absolute bias to less than 10%.

Recognizing the left-truncation nature of prevalent cases, the current analysis excluded prevalent HPV DNA detected at baseline. Women may have experienced one or multiple episodes of gain or loss of HPV DNA detection before they were recruited and, therefore, the first incident detection (or loss of detection) under observation may simply be one of the many recurrent events. As such, we included the first incidence along with other recurring events of detection (loss of detection) in the following survival analysis. For detection events, the time of origin began at the mid-interval of a positive type-specific test and a previous negative result; the timing of an event was similarly calculated as the mid-interval of the positive visit and a subsequent visit with a negative test. Likewise, for recurrent events of loss of detection of a specific type, women began to enter the risk set at the mid-interval of a positive test and the following negative test. Although recurrent events of detection and loss of detection for all 37 types within the same individual were observed over the study period, event history analysis methods were used for these repeated events (12). Survival functions were estimated by the generalized product limit estimator developed by Peña and colleagues to account for correlation among multiple events of the same genotype over time as well as a potential nonindependence among different genotypes within the same woman (13, 14). Stratum-specific survival functions were also determined by stratifying on potential risk factors. Woman-level risk factors included older age (>40 years), number of lifetime sex partners, and self-reported casual sex history at baseline. Self-reported sexual activity at follow-up, such as vaginal intercourse and condom use, were summarized as time-invariant covariates at the end of the study. All descriptive analyses were performed using Stata for Windows version 11.2 and the “survrec” package (version 1.2-2) of the R statistical program, version 3.0.0 at a two-sided significance level of 0.05 unless otherwise specified (15, 16).

Over the study period, 28 of 33 women completing both phases I and II of the original study were positive for at least one type of HPV DNA (Table 1). On average, each of these 28 women contributed a median number of 30 mid-vaginal specimens [inter-quartile range (IQR), 29–31; total: 955], recorded four menstrual cycles (median, 4; IQR, 4–5) and kept 113 daily diary records (median; range, 112–115). There was little evidence of risk of new exposures during the 16-week period for these 33 women. Most women reported being in a monogamous relationship. The only exception was 3 women (9%) reporting ever having casual partners in the prior 12 months at baseline, and 1 (3%) reported a new partnership at follow-up (Table 1).

The temporal dynamics of detectable viral DNA is shown in Fig. 1 (also see Supplementary Figs. S1 and S2). Overall, a median number of four unique HPV genotypes (IQR, 3–7; range, 1–13) were detected per ever-positive woman (Fig. 1). In comparison, the median number of HPV types detected on one randomly collected swab was two (IQR, 3–4). Within the same subject, an ever-positive genotype was only detectable in 49.3% [95% confidence interval (CI), 23.7%–90.3%] of all swabs that were positive for at least one HPV type, suggesting that single-time swab results have a low-to-intermediate reliability for type-specific HPV DNA detection.

Table 2 shows comparisons between point and cumulative prevalence of detectable viral DNA. At baseline, 20 women had at least one genotype of HPV DNA detected in the first available swab, corresponding to a point prevalence of 60.6%. The 16-week cumulative prevalence was higher at 84.8%. Assuming that cumulative prevalence represented the “true” underlying infection status, the absolute bias comparing point versus cumulative prevalence was −24.2% and the relative bias was −28.6% (Table 2). Similarly, the prevalence of high-risk types was underestimated by 36.4% (absolute) and 60.0% (relative), respectively. Results of pair-wise exact McNemar tests were all statistically significant, indicating that the point prevalence estimate was substantially different from the corresponding cumulative prevalence. Bootstrapped results provided a wide range of possible biases when the “baseline” sample used for point prevalence estimation was randomly sampled over the 16-week period. In general, point prevalence underestimated cumulative prevalence by 20% to 36% in absolute values and by 24% to 55% in relative terms (Table 2). To reduce the median absolute bias in point prevalence of any-type HPV to less than −10%, four or more consecutive swabs would be needed given the cumulative prevalence in the current study and the sampling interval used (Table 2). Similar results to reduce biases when requiring nonconsecutive sampling are provided online (see Supplementary Table S1).

To assess the robustness of our results, we excluded genotypes that appeared only once (“sporadic detections”) for each woman over the study period and repeated analyses for the bias quantification. Among 136 ever-positive person-types, 29% were positive on single occasion (n = 40), whereas 96 person-types (71%) were detected at least twice (see Supplementary Table S2). After excluding these “sporadic” detections, the cumulative prevalence of any-type HPV or of high-risk types remained the same. The only difference observed was a lower cumulative prevalence of women with multiple-type HPV infection than previously estimated (Table 2), but this difference was not statistically significant (63.6% vs. 66.7%; P = 1.000 for exact McNemar test). As a result, empirical or bootstrapped estimates for misclassification bias were overlapping with those based on all ever-positive results (data not shown).

During follow-up, 319 type-specific viral detections were recorded among 27 women (768 person-swabs; Table 3). One woman was persistently positive for the same type over the entire study period. Overall, each woman experienced a median number of 10 incident events (IQR, 5–14; Table 3). The estimated median time to a new detection was 11 days (95% CI, 10–11). Women with 5 or more lifetime sex partners experienced a 3-day shorter interval between any two events than those without (11 vs. 14 days), suggesting that more recurrent events (median: 10 vs. 9 events) were marginally associated with a higher cumulative sexual exposure. Similarly, women reporting casual sex partners in the previous 12 months were more likely to have frequent incident detections than those without (median time to incident detection: 7 vs. 11 days). Older age (>40 years) did not seem to affect women's risk of HPV DNA detection (Table 3).

Meanwhile, 313 losses of type-specific detection were documented among these 27 women; the median number of loss-of-detection events was 10 (IQR, 5–14; Table 4). The median time to a recurrent loss of detection was 7 days and varied little across strata of selected woman-level covariates, except for history of casual sex reported at baseline. Women with a prior exposure to casual sex tended to have a marginally longer detectable duration for HPV DNA than those without [median (IQR): 7 (4–7) vs. 10 (7–14) days; Table 4]. For recurrent detection and loss of detection, we did not find significant associations between these recurring events and any sexual intercourse recorded during the sampling interval.

We analyzed 955 vaginal swabs from 33 predominately monogamous women over a 16-week period and revealed substantial variation in detectable HPV DNA over time. Consistent with the growing body of literature, we found that point prevalence estimates generally underestimate cumulative prevalence by at least 20% (17–22). Under an assumption that the variability reflects gain and loss of prevalent DNA detection among infected individuals, we estimated that at least four consecutive swabs per woman would be required to minimize the (absolute) median bias of “true” infection status to less than 10%. However, even if the assumption of variable detection in truly infected individuals is valid, we recognize that collection and testing of multiple samples during each sampling interval is not feasible in large natural history studies. This analysis was therefore conducted to quantify possible misclassification in infection status to aid in interpretation of long-interval sampled data, and not to suggest new sampling algorithms.

An alternative explanation for the variability in repeated HPV testing results is that these variations represent new acquisitions, from either transient depositions of viral material from a recent sex act (23), or reinfections of the same genotype (5, 24). In this study, most women were in a monogamous relationship during the study period, and thus had a relatively low risk for new infections. Although frequent reinfection and then clearance from the single male partner could explain the observed variability, we found that the temporal changes were no less frequent among abstinent women than those who were sexually active throughout the follow-up period. Yet, like many others, we were unable to make a clear distinction between persistent infections of low viral load fluctuating around the assay detection limit and reinfections of the same type. Even if women were conferred nonignorable risks of reinfection by their partners' behavior (about which we had no information), women would also have to be susceptible to the same type that they were infected with as recently as less than 2 weeks ago. However, according to animal models, it generally requires 2 to 3 weeks for HPV to establish a successful life cycle in tropic tissue cells (25). Because of our uncertainty about the normal duration of natural immunity following a transient infection, we cannot definitively rule out the possibility of repeated acquisition and clearance events with frequent exposure to an infected male partner. Nevertheless, the short interval between two new detections found in the current study (1–6 weeks, type-dependent) reduces the probability of reinfection as the only probable cause of the variable HPV detection.

As a result, we considered low-level persistent infections or intermittent reactivation from viral latency as an equally probable cause of the observed variability in HPV detection. This explanation would be consistent with our observation that women who remained abstinent over the study period and were thus not at risk for new acquisition or reinfection had similarly high variations in viral detection (and nondetection). We have also recently found that, using 6-month interval sampling, risk of new HPV detection in peri-menopausal women was similar in sexually abstinent and monogamous women, who contributed 85% of all new HPV detection (26). In that study, the HPV population attributable risk (PAR) due to a higher number of lifetime sex partners (PAR = 72%) was higher than that associated with recent new sex partners (PAR = 13%–27%), consistent with variable detection of low-level persistent infection or reactivation from latency (26). These observations are further supported by prior work showing that detecting a previously found, type-specific HPV DNA following a period of nondetection is often a result of redetection of the same HPV variant rather than reinfection of the same genotype even among sexually active women (1, 4, 27).

Several limitations are noted while interpreting our findings. First, HPV prevalence estimates were based on mid-vaginal swabs using sampling devices and DNA extraction methods were not typically used in HPV natural studies, which may raise concerns about the representativeness of the more clinically relevant cervical infections and probability of sampling errors. However, a systematic review published in 2005 has concluded the agreement of HPV prevalence as determined by self-collected vaginal specimens with that by physician-collected cervical samples (28). This statement was further supported by two more recent reviews (29, 30). Second, we have ignored the potential phase dependence while estimating bias distributions (31, 32), and our bootstrapping approach may have introduced additional sampling errors while performing resampling. Third, lack of statistical power because of the highly correlated nature of multiple events per subject was another concern in interpreting stratified results of the current analysis. For the same reason, we were unable to assess other potential time-dependent factors that might influence HPV detection, such as the composition of the vaginal bacterial communities (33). Finally, our duration estimates should be interpreted as duration of incident detection events, and not as overall duration of HPV infection as Weaver and colleagues have shown that estimated durations of low-level persistent infections using standard assays are likely to be severely underestimated (4).

In summary, we found that HPV DNA detection was extremely variable over very short intervals. We quantified biases in point estimates as compared with cumulative prevalence when assuming the variability reflected changes in detectability rather than in infection status. We reasoned that, in addition to new acquisitions or reinfection of the same type via sexual exposures, redetection of low-copy persistent infections or reactivated latent infections seemed equally plausible in explaining the dynamics of variable HPV DNA detection.

D.A.T. Cummings has commercial research grant from Medimmune and is a consultant/advisory board member of the same. No potential conflicts of interest were disclosed by the other authors.

Conception and design: S.H. Liu, D.A.T. Cummings, P.E. Gravitt, R.M. Brotman

Development of methodology: S.H. Liu, D.A.T. Cummings, J.M. Zenilman, P.E. Gravitt

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.M. Zenilman, P.E. Gravitt, R.M. Brotman

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.H. Liu, D.A.T. Cummings, P.E. Gravitt, R.M. Brotman

Writing, review, and/or revision of the manuscript: S.H. Liu, D.A.T. Cummings, J.M. Zenilman, P.E. Gravitt, R.M. Brotman

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.H. Liu, P.E. Gravitt, R.M. Brotman

Study supervision: P.E. Gravitt, R.M. Brotman

The authors thank Roslyn Howard for her laboratory work and the participants of the douching cessation study for their voluntary involvement and generous devotion of time and efforts. The authors also thank Dr. Gary Rosner at The Sidney Kimmel Comprehensive Cancer Center, Johns Hopkins University, for his insightful comments on statistical methods.

This work was supported by the U.S. National Institute of Allergy and Infectious Diseases (K01 AI080974 to R.M. Brotman and R03 AI061131 to J.M. Zenilman).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.