Clinical Utility of Reflex Testing with Cancer Biomarkers to Improve Diagnostic Accuracy of Primary Human Papillomavirus Screening

Human papillomavirus (HPV) testing is now considered the optimal, most efficient screening approach for cervical cancer in both low- and high-resource settings (1–6). Compared with cytology, HPV testing has proven to be considerably more sensitive in detecting cervical intraepithelial neoplasia (7). A single round of HPV testing provides greater and lengthier reassurance against future risk of developing precancer/cancer and avoids the frequent, repeated testing associated with cytology (8, 9). Furthermore, the availability of low-cost, point-of-care HPV tests allows for screening and treatment to occur in a single visit, which greatly reduces opportunities for loss-to-follow-up (10).

The main disadvantage of HPV testing is its relatively low specificity arising from its inability to fully discriminate between transient and persistent infections (5, 6, 11). It is well established that persistent infection with one of the 14 high-risk HPV genotypes plays a central role in cervical carcinogenesis (12, 13). However, only a subset of women with high-risk HPV progress to develop invasive cervical cancer because most infections will spontaneously clear (5, 14, 15). This challenge is particularly marked in women living with HIV who have high rates of HPV infections (16). To reduce overtreatment in the screen-and-treat model, there is a need to better identify HPV infections at highest risk of evolving into invasive cervical cancers and thereby improve risk stratification of women within screening programs (6, 17, 18). For women living with and without HIV, we have previously evaluated two approaches to improving diagnostic accuracy of a single round of HPV testing. We have shown improved specificity with minimal reduction in sensitivity through: (i) cycle threshold optimization—to restrict to more stringent cycle threshold (Ct) cutoff values (higher viral loads); and (ii) genotype selection—to select a subset of highest-risk HPV genotypes (19, 20).

In this analysis, we aimed to determine whether inclusion of cancer biomarkers as reflex tests following a positive HPV result can further improve the diagnostic accuracy of a round of cervical cancer screening. We selected the following three biomarkers from those that have shown promise in accurately identifying women with precancerous cervical lesions: cyclin-dependent kinase inhibitor 2A (CDKN2A), topoisomerase 2 alpha (TOP2A), and proliferation marker Ki67 (Mki67; refs. 18, 21–23). CDKN2A is a tumor suppressor gene that encodes for p16 and prevents cell-cycle progression through the retinoblastoma (pRB) pathway (23–26). In normal functioning cells, p16 inactivates cyclin-dependent kinases that phosphorylate pRB and causes the cell to slow its progression from G₁ phase to S phase. In transforming cells infected with high-risk HPV, the oncoprotein E7 degrades pRB causing the cell to proceed with division and p16 is overexpressed in attempt to inhibit the abnormal proliferation. TOP2A encodes for an enzyme involved in chromosome condensing, chromatid separation and unwinding of DNA strands during S phase (23, 27). Although the role of Mki67 is less clear, it is required for cell division and present in all phases of the cell cycle except G₀ (28). TOP2A and MKi67 are also overexpressed when cells aberrantly enter cell division.

In a clinical study in Cape Town, South Africa, we investigated whether the addition of reflex testing for CDKN2A, TOP2A, and/or Mki67 following HPV testing can improve the diagnostic accuracy of a round of cervical cancer screening as compared with our previous approaches of genotype selection and cycle threshold optimization (19, 20).

We recruited 1,121 women ages 30–65 years from two sites in Cape Town, South Africa. A community-based clinic in Khayelitsha was used to recruit women seeking routine women's health services and was considered representative of a healthy, screening population. A colposcopy clinic in Groote Schuur Hospital where women were referred for abnormal cervical cytology was used to obtain a referral population. Women were eligible to participate in the study if they had a documented HIV status, no history of treatment for cervical dysplasia, no previous hysterectomy, and no current pregnancy. An approximately equal number of women living with and without HIV were recruited by design. The study was approved by institutional review boards at the University of Cape Town and Columbia University. Written informed consent was obtained from each participant.

During pelvic examination, a physician collected two cervical samples from each woman in separate 20 mL vials of PreservCyt. One sample was tested at point-of-care using Xpert HPV, which uses real-time PCR to detect 14 high-risk HPV genotypes. Xpert HPV reports cycle threshold (Ct) values for five separate channels: HPV16, HPV18, 45, HPV31, 33, 35, 52, 58, HPV51, 59, and HPV39, 56, 66, 68. For each channel, the manufacturer provides predetermined cutoff Ct values that define a positive HPV result (40 cycles for HPV channels 16 and 18, 45; 38 cycles for the other 3 channels).

For the first 1,000 participants, the second cervical sample was sent to the Cepheid Research laboratory in Sweden to test for three biomarkers: CDKN2A, TOP2A and the Mki67 using PCR.

Each PreservCyt sample (5 mL) was used for RNA preparation. The samples were centrifuged at 300 × g for 5 minutes, the cell pellets were mixed with 700 μL QIAzol (Qiagen) lysis reagent and stored in −80°C until extraction. Total RNA was extracted using the RNeasy RNA extraction kit (Qiagen), according to the manufacturer's protocol. RNA concentration was measured with a NanoDrop instrument (Thermo Scientific). Reverse transcription was performed starting with 10 μL of total RNA in a 20 μL reaction volume using random hexamers and the High Capacity cDNA RT kit from Applied Biosystems Inc according to the manufacturer's protocol.

All PCR reactions were performed in duplicate in a reaction volume of 25 μL containing 5 μL of cDNA, diluted 1:10, mixed with TaqMan Universal PCR MasterMix (Applied Biosystems Inc). The following protocol was used for all assays: denaturation (95°C for 10 minutes) and amplification (95°C for 20 seconds, 60°C for 1 minute) repeated for 40 cycles. The expression of CDKN2A, TOP2A, and MKI67 was analyzed. Beta glucuronidase (GUSB) and cGMP-dependent protein kinase 1 (PRKG1) were used as reference genes for normalization; these genes have previously been demonstrated to have a high stability in expression between groups of normal samples versus HSIL samples (18). Signal cutoff for target genes was set at 40 cycles, whereas for the reference gene PKG1, a Ct value above 30 cycles indicates poor RNA quality and was therefore considered invalid and excluded from the analysis. Primers and probes used in the PCR are listed in Supplementary Table S1. All probes were FAM-labeled and all reactions were run in single-plex.

After collection of cervical samples, the physician conducted colposcopy to obtain histologic samples by biopsy or loop electrosurgical excision procedure for women with visible abnormal-appearing areas of the cervix or by endocervical curettage for women with no visible lesions. Second and third colposcopies were conducted for women with positive HPV results at the screening clinic who had not had cervical cancer precursor lesions detected by histologic sampling at the first colposcopy. Random four quadrant biopsies were done at the repeat colposcopies to reduce the likelihood of missing disease. Histologic samples were reviewed initially by a pathology laboratory in Cape Town and then sent to the USA for blind rereview to determine cervical intraepithelial neoplasia grade 2 or worse (CIN2⁺) and grade 3 or worse (CIN3⁺). If needed, a third US pathologist was used to resolve any discrepancies.

All statistical analyses were conducted using SAS software version 9.4.

Biomarkers were treated as continuous variables with the intensity of the expression of the biomarker quantified as delta cycle threshold values (ΔCt = Ct _biomarker − Ct _{reference gene}). Because higher Ct values indicate lower expression of the biomarker, higher ΔCt values indicate lower expression of the biomarker relative to the reference. We used the following parameters to characterize the distributions of each biomarker by HIV status and cervical disease status: median, 25th percentile, 75th percentile, and range. To assess each biomarker's ability to discriminate women with and without cervical disease, we used Kruskal–Wallis tests to compare biomarker distributions between women with <CIN2 and CIN2⁺ by HIV status. We further conducted logistic regression analysis to determine if the three biomarkers, alone or in combination, predicted CIN2⁺ or CIN3⁺. Logistic regression models were run separately for women living with and without HIV. Of note, women from both the screening and referral populations were combined to create the disease endpoint group of CIN2⁺ or CIN3⁺. Only women from the screening population were included in the no-disease endpoint group (<CIN2 or <CIN3).

We defined HPV positivity in two different ways: (i) positive for any of the five channels detecting the 14 high-risk HPV types and (ii) positive only for three select channels detecting HPV16, 18, 45, 31, 33, 35, 52, and 58. Both definitions of HPV positivity utilize the manufacturer-defined Ct values. We calculated sensitivity and specificity of HPV testing using conventional methods. Sensitivity was calculated using the screening and referral populations, whereas specificity was calculated with the screening population only.

We used multivariable logistic regression models to evaluate the utility of biomarkers as reflex tests after HPV testing. Each of the three biomarkers as reflex tests was entered individually into the regression models, as well as together, and those that remained statistically significant (P < 0.05) were retained in the final multivariable model. When the HPV test was positive, we reversed the direction of the delta Ct values for each biomarker (12 − ΔCt). When the HPV test was negative, we coded the biomarker as 0.

Furthermore, we compared logistic regression models for biomarkers as reflex tests to the alternative triage strategy of cycle threshold optimization. As previously described in Kuhn and colleagues, Ct values from the three informative channels on the Xpert assay were utilized in logistic regression models (19). In brief, channel-specific reverse cycle threshold (rCt) values (45 − Ct) for the three channels detecting HPV types 16, 18, 45, 31, 33, 35, 52, and 58 were used when the manufacturer's cutoff defined the test result as positive. If the result was defined as HPV negative based on the manufacturer's cutoff, then the rCt was equal to 0.

From each of the logistic regression models, we generated receiver operating characteristic (ROC) curves and the area under the curves (AUC): 0.7–0.8 acceptable, 0.8–0.9 excellent, 0.9–1.0 outstanding (29). 95% Confidence intervals (CI) were calculated for the AUC estimates. We selected a sensitivity of 80% as the benchmark for the cutoffs on biomarkers and the HPV viral load measures, which we reasoned was adequate for a round of screening. We derived corresponding specificities at 80% sensitivity using the ROC curves from each logistic regression model. In addition, we calculated the positive and negative predictive values and the proportion who would test positive at the identified sensitivity/specificity combinations. For these parameters, calculations were based on the observed prevalence of CIN2⁺ or CIN3⁺ in the screening population and standard algebraic formulae derived from the 2×2 table. Delta Ct cutoff values for the biomarkers were calculated using the probability derived from the ROC curve at sensitivity of 80% and the beta and intercept parameters from the logistic regression model. Formulae are available in Supplementary Materials and Methods.

One thousand of the 1,121 women recruited in the original study were additionally tested for the biomarkers CDKN2A, TOP2A, and MKi67. Demographic characteristics of this study population have been described elsewhere (19). Six-hundred sixty-eight women (355 without HIV, 313 with HIV) were from the screening population whereas the remaining 332 women (130 without HIV, 202 with HIV) were from the referral population. After exclusions for missing HPV results (n = 1), invalid biomarker data (n = 8), missing and inadequate pathology results (n = 12), there were 251 women with CIN2⁺ (85 without HIV and 166 with HIV) from the screening and referral populations combined. For the nondisease group, only women in the screening population were included. There were 588 women with <CIN2 (331 without HIV and 257 with HIV; Supplementary Fig. S1).

The prevalence of confirmed cervical intraepithelial neoplasia grade 2 or worse (CIN2⁺) was 5.16% (95% CI, 2.84–7.48) and 16.29% (95% CI, 12.16–20.42), and the prevalence of high-risk HPV 16.62% (95% CI, 12.71–20.52) and 48.21% (95% CI, 42.62–53.80) among women without and with HIV, respectively, in the screening population (Table 1).

Table 1 shows the conventional sensitivity and specificity of HPV testing. Sensitivity of Xpert HPV to detect CIN2⁺ was 88.24% (95% CI, 81.39–95.08) and 93.37% (95% CI, 89.59–97.16) among women living without and with HIV, respectively, when defining screen positive as detection of any the 14 high-risk HPV types (positive on any of the five channels). Specificity of HPV testing, with the same definition of screen positive, was 86.40% (95% CI, 82.71–90.10) and 59.92% (95% CI, 53.93–65.91) among women living without and with HIV, respectively.

Shifting to defining screen positive as detection of HPV16, 18, 45, 31, 33, 35, 52, and/or 58 (positive on only three of the channels) had little reduction in sensitivity. However, specificity improved to 89.43% (86.11–92.74) and 66.93% (61.17–72.68) in women living without and with HIV, respectively (Table 1). The high sensitivity but relatively low specificity of HPV testing, particularly in women living with HIV, was the prime justification for investigating additional approaches. In a round of screening, our goal was to determine how much improvement we could achieve in specificity when allowing sensitivity to decline to a threshold of 80%.

The distributions of CDKN2A, TOP2A, and MKi67 for women with CIN2⁺ were distinct from that of women without CIN2⁺ (Fig. 1). For each biomarker, women with CIN2⁺ had higher expression of the biomarkers (lower ΔCt values) than women without CIN2⁺. For women without HIV, the median ΔCt values were 0.80 versus 2.38 with CDKN2A, 2.60 versus 5.43 with MKi67, and 3.13 versus 5.99 with TOP2A for CIN2⁺ and <CIN2, respectively. For women with HIV, median ΔCt values were 1.12 versus 2.69 with CDKN2A, 3.61 versus 5.40 with MKi67, and 3.94 versus 5.60 with TOP2A for CIN2⁺ and <CIN2, respectively (Fig. 1, P < 0.001 for all comparisons). Among women without CIN2⁺, the distribution of the biomarkers did not differ by HPV status. Among women with CIN2⁺, the distribution of the biomarkers was shifted toward less expression in women with HPV among women without HIV only (Supplementary Table S2).

With reflex testing, biomarkers are used to further triage women who have already tested positive for HPV. First, we considered using the HPV cutoffs defined by the manufacturer to detect all 14 high-risk types on all five channels as the basis to prompt a biomarker reflex test; and second, we considered the results if the definition of a positive result was restricted to the three channels that detect HPV types 16, 18, 45, 31, 33, 35, 52, and 58. In univariate regression models, each biomarker was significantly associated with CIN2⁺ among women who tested HPV positive using either definition. When included together in multivariable regression models, TOP2A was no longer significant. Both CDKN2A and MKi67 remained significantly associated with CIN2⁺.

Best results were obtained for models that included ΔCt values for both CDKN2A and MKi67. Among those who tested positive for any of the 14 high-risk types, the AUC using the two biomarkers together to predict CIN2⁺ was 0.9013 (95% CI, 0.8494–0.9533) for women without HIV and 0.7993 (95% CI, 0.7447–0.8539) for women living with HIV. Results were largely similar among women with the genotype-restricted definition of a positive HPV test where the AUC using the two biomarkers together to predict CIN2⁺ was 0.8988 (95% CI, 0.8427–0.9549) for women without HIV and 0.7949 (95% CI, 0.7358–0.8540) for women living with HIV. Performance of biomarkers was consistently better (higher AUC values) in women living without HIV compared with women living with HIV (Fig. 2).

Table 2,³ displays the performance characteristics of the biomarkers when undertaken as reflex tests (i) among women who tested positive for any of the 14 high-risk types or (ii) among women who tested positive for the selected high-risk HPV types. Utilizing the ROC curves from the logistic regression models, we determined the specificity that would be attained if sensitivity of a round of reflex testing was set at 80%. We also determined the ΔCt values for each biomarker that would be needed for a woman to be considered screen-positive and achieve a sensitivity of 80%. We further calculated positive and negative predictive values and percentage of women who would be screen positive at these cutoffs for ΔCt values.

Each biomarker alone was an effective triage after a positive HPV test. However, the combination of CDKN2A and MKi67 yielded the best results for women living with and without HIV. Benchmarking sensitivity at 80%, specificity for a round of reflex testing with CDKN2A and MKi67 after a positive HPV test was 95.17% in women without HIV and 81.32% in women with HIV. Slightly better specificity for a round of reflex testing with the biomarkers were observed in women with HIV if the definition of a positive HPV test was confined to the subset of high-risk types. Results also favored using the selective HPV types as the initial screen as this reduced the proportions who would need a second test and thus the overall proportions defined as screen-positive (Table 2).

We have previously shown that the sensitivity/specificity trade-off of HPV testing can be improved by shifting thresholds for the cutoff values on the three Xpert channels detecting HPV16, 18, 45, and 31, 33, 35, 52, 58 (19). As shown in Fig. 3, AUC was excellent for women without HIV (AUC = 0.9059) and women living with HIV (AUC = 0.8705) when HPV results were treated quantitatively using the reverse Ct values on the three informative HPV channels. When sensitivity was set at 80%, utilizing the HPV Ct values alone led to specificity rates of 93.96% and 83.27% in women living without and with HIV, respectively. These performance characteristics were near identical to those achieved with reflex testing with the biomarkers. Results using CIN3⁺ as the endpoint produced essentially similar results (Fig. 3).

HPV testing is the optimal, most efficient screening approach for both low- and high-resource settings and has the highest sensitivity and negative predictive values of any known approach to cervical cancer screening (1–6). Despite these strengths, there has been some reluctance to embrace HPV testing for screening in low-resource settings because of its relatively low specificity, particularly for women living with HIV. However, the very high sensitivity of HPV testing, as it is currently configured, is arguably not necessary for most screening programs. From a public health standpoint, a degree of compromise in sensitivity, if the concomitant gain in specificity is substantial enough, would lead to more effective screening programs. This logic has prompted much interest in cancer biomarkers that could be used in conjunction with HPV testing to help better triage women who test positive on HPV tests (18, 26).

In our study, each of the three biomarkers we evaluated, CDKN2A, TOP2A, and MKi67, strongly discriminated between women with and without CIN2⁺. If used as triage tests following HPV testing, cutoffs for the expression of these biomarkers can be identified that would result in favorable performance parameters for a round of screening. Consistent with prior studies (18), CDKN2A was the most specific biomarker, likely given its expression as a direct consequence of transforming HPV viral oncoproteins E6 and E7 (26). In contrast, TOP2A and MKi67 are nonspecific markers of cell proliferation (26). In our study, the combined expression of CDKN2A and MKi67 led to the best differentiation between women with and without CIN2⁺. One of the strengths of our study methodology is the use of molecular testing on liquid-based samples to detect mRNA through PCR (18). Several prior biomarker analyses have been done with IHC, which risks intraobserver subjectivity when conducting microscopic analysis of histologic samples.

From the same study (19), we have previously reported our results pertaining to the improvements in the sensitivity/specificity balance that can be achieved by shifting the cycle threshold values on the three informative channels of Xpert HPV that detect HPV types HPV16, 18, 45, 31, 33, 35, 52, and/or 58. Here we observed that the magnitude of the benefit achieved with the inclusion of the cancer biomarkers did not exceed that achieved with the changes to the interpretation of the output of the Xpert HPV assay that we have previously evaluated. Several groups are evaluating biomarkers as triage tests for women who have positive HPV results. Our results provide proof of concept that this is a viable approach, should appropriate tests become commercially available and affordable. Nevertheless, inclusion of biomarkers as reflex tests has several disadvantages, particularly for low-resource settings. Notably, the costs can be prohibitive for an additional test in those who require it as well as additional logistic steps and delay before reaching a clinical decision required in this approach. Xpert HPV is currently formulated to be easily used at the point of care allowing for single-visit screen-and-treat approaches to be implemented.

New World Health Organization guidelines propose the need for triage tests after HPV testing for primary screening, particularly for women living with HIV who have high HPV prevalence rates. Our study was designed to consider the performance of screening modalities in women both living with and without HIV. This is particularly relevant for a setting like South Africa with extraordinarily high rates of HIV infection among women. Our study was also designed for a setting where resources are scarce and where screening programs have failed to be effective due to attrition inherent in the multistep cascade of conventional screening algorithms. Given the high prevalence of HPV among women living with HIV, our approach has marked benefit for these women and would substantially reduce overtreatment in the screen-and-treat model. For women without HIV, HPV testing without any further triage has adequate performance parameters, but the adaptation to the interpretation of the output of the HPV test that we have evaluated optimizes this further.

Although rigorously designed, there are several limitations to our study. Due to limited resources, we used an “enrichment method” to recruit women from both a screening population and a referral population to attain sufficient numbers of women with cervical disease. Therefore, the study sample may not be representative of disease detected in a true screening population and the sample size was too limited to investigate directly the generalizability here. This enrichment method was utilized and supported in the VALGENT HPV test validation (30).

In conclusion, our results provide proof of principle that biomarkers can be effectively used as a triage after positive HPV results to improve the overall sensitivity/specificity balance in a round of cervical cancer screening. Excellent performance parameters can be attained even for women living with HIV. However, addition of reflex testing for biomarkers does not further improve on the modifications to interpretation of the Xpert HPV assay that we have previously evaluated: cycle threshold optimization and genotype restriction (19). For settings without access to HPV assays that have characteristics similar to Xpert HPV, namely, the capacity to obtain partial typing and markers of viral quantitation, biomarker assays will have clinical utility. For settings with access to Xpert HPV or similar assays, use of a single HPV assay that can be done at the point of care is more desirable from a programmatic standpoint.

L.G. Johnson reports grants from NCI during the conduct of the study. R. Saidu reports grants, personal fees, and nonfinancial support from NIH during the conduct of the study. J. Moodley reports grants from NIH during the conduct of the study. A. Tergas reports other support from Auro Vaccines and Immunomic Therapeutics outside the submitted work. D. Persing reports personal fees from Danaher Corporation during the conduct of the study. S.A. Campbell reports personal fees from Cepheid during the conduct of the study as well as personal fees from Cepheid outside the submitted work. T.C. Wright Jr reports personal fees from Roche and Becton Dickerson during the conduct of the study. No disclosures were reported by the other authors.

L.G. Johnson: Formal analysis, visualization, writing–original draft, writing–review and editing. R. Saidu: Investigation, writing–review and editing, recruitment; clinical follow-up. C. Svanholm-Barrie: Formal analysis, investigation, writing–review and editing, biomarker analysis. R. Boa: Investigation, writing–review and editing, recruitment, clinical follow-up. J. Moodley: writing–review and editing. A. Tergas: Writing–review and editing. D. Persing: Investigation, technical assistance on human papillomavirus testing with Xpert. S.A. Campbell: Investigation, writing–review and editing, technical assistance on human papillomavirus testing with Xpert. W.-Y. Tsai: Formal analysis, writing–review and editing. T.C. Wright Jr: writing–review and editing. L. Denny: Conceptualization, supervision, funding acquisition, investigation, writing–review and editing. L. Kuhn: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

The study was supported in part by the NCI UH2/3 CA189908. L.G. Johnson is supported by the NCI T32 Postdoctoral Fellowship in Cancer Epidemiology at Columbia University Mailman School of Public Health (T32CA094061). We would like to thank Dr. Gwynn Stevens, Ms. Dipti Lallubhai, and Ms. Tessa Visser of Cepheid–South Africa; Ms. Jill Birkmeier and the Cepheid Clinical Affairs and Data Management and Analytics team for study record capture, database construction, and cleaning; and the facility managers and senior leadership of the Department of Health, Western Cape Provincial Government for support with the study.

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.