As a group, human papillomaviruses (HPV) are proven human carcinogens. But there are >> 100 HPV genotypes and only a small fraction have any known carcinogenic potential [1]. Therefore, moving from broad acceptance of the carcinogenicity of persistent HPV infection to specific conclusions about individual genotypes requires consideration of each type as an individual agent. Such a type-by-type evaluation proves to be very difficult, and stretches epidemiology to its limits because of issues of confounding and misclassification detailed below. However, as described in the accompanying article by Castle [2], it is important for the development of screening tests and vaccines to judge each HPV type separately; thus, it is worth considering how well one actually can decide whether a given HPV type is carcinogenic or not. This article describes an approach taken by a Working Group of the World Health Organization International Agency for Research on Cancer (IARC), when asked to re-assess for official purposes which HPV types should be grouped as carcinogens [3]. The process is described here to promote needed debate on what an improved approach might be.

IARC formally considers the carcinogenicity of exposures to humans. Whereas human carcinogenicity might best be considered for some agents like HPV as a continuum of probabilities without a clear breaking point, IARC classifies carcinogens categorically as carcinogenic (Group 1), probably carcinogenic (Group 2a), possibly carcinogenic (Group 2b), not classifiable (Group 3), or probably not carcinogenic (Group 4). There has been very little experimental work on the carcinogenicity of HPV types except for HPV16 and HPV18; thus, epidemiologic evidence has been unusually important. Epidemiologic evidence for the carcinogenicity of HPV was originally presented in IARC Monograph Volume 64 [4], and was extensively updated in IARC Monograph Volume 90 based on data available as of February 2005 [5]. In February 2009, Volume 100b updated the data once again and it is this latest update addressed here. For the purposes of IARC, epidemiologic evidence is categorized as sufficient, limited, inadequate or suggesting lack of carcinogenicity. The epidemiologic data are combined with experimental evidence (in this case lacking for most HPV genotypes) to arrive at the final Groups 1–4.

HPV carcinogenicity has been established most convincingly for cervical cancer, and this discussion will be limited to cervical carcinogenicity. To date, no HPV type has been proven to be carcinogenic only at sites other than the cervix.

HPV behaviour at the cervix is strongly correlated with phylogenetic (i.e. evolutionary) categories [6]. All HPV genotypes that are known to be cervical carcinogens belong to the alpha genus in an evolutionary branching or clade containing a few genetically related species (Figure 1). Epidemiologic data do not support cervical carcinogenicity for other species in the alpha genus or for other genera. To save considerable space presenting null evidence, this section will not include data related to HPV species alpha-1, -2, -3, -4, -8, -10 (other than HPV 6), -13, or 14/15. These species contain HPV types that cause skin or genital warts, minor cytologic atypia, and often no apparent disease.

In Monograph 64 in 1995, HPV 16 (alpha-9) and HPV 18 (alpha-7) were classified as cervical carcinogens. HPV 31 and HPV 33 in alpha-9 were categorized as probably carcinogenic [4]. In 2005, the group of cervical carcinogens (Group 1) was expanded to include the following 13 types: alpha-5 genotype HPV 51; alpha-6 genotypes HPV 56 and HPV 66; alpha-7 genotypes HPV 18, HPV 39, HPV 45, and HPV 59; and alpha-9 genotypes HPV 16, HPV 31, HPV 33, HPV 35, HPV 52, and HPV 58 [5].

In the four years between Monograph 90 in 2005 and the recent update, new evidence further supported that HPV types in the high-risk clade of the alpha genus HPV type cause virtually all cases of cervical cancer worldwide [7, 8]. In case-control studies, the odds ratios associating cervical cancer and its immediate precursor, CIN3, with HPV DNA positivity for these types in aggregate has consistently exceeded 50. It is persistent infections that are associated with extremely high absolute risk of CIN3 and cancer. In cohort studies, women who test negative to this group of HPV types are at extremely low subsequent risk of CIN3, cancer, and cancer death for more than 10 years [9–11].

From a virologic perspective, the definitive proof of carcinogenicity of an HPV type is finding transcriptionally active HPV in a tumour. HPV is not a "hit and run" carcinogen, and transcriptional activity is needed for maintenance of the cancer phenotype. In cervical cancer cell lines, blocking transcriptional activity by antisense RNA leads to apoptosis. This level of evidence is simply lacking for virtually all HPV types. The vast body of evidence relates simply to finding HPV DNA at the same time of cervical neoplasia.

However, relying on testing of scrapes or biopsies, by DNA testing, leads to difficulties. The alpha HPV types share a common route of transmission and multiple infections are present in a large minority of women, although they might not be transmitted from the same partner or at the same time. Given the existence of some very powerfully carcinogenic types, notably HPV 16 and HPV 18, determining which weaker and/or less common types are also carcinogenic becomes (for the epidemiologist) an issue of confounding. None of the traditional approaches to control of confounding are entirely successful. Because HPV 16 causes approximately 50% of cases of cervical cancer, logistic regression and similar approaches will parsimoniously attribute cases associated with both HPV16 and a less important type to HPV16. HPV 18 is the second most important cervical carcinogen, responsible for approximately 15–20% of cervical cancer of all histologic types combined (and a higher fraction of adenocarcinomas). If a type occurs with either HPV 16 or HPV 18, its association with cervical cancer might be confounded by either of these powerful carcinogens. For types causing only a very small fraction of cervical cancer, confounding by any of the more important types is possible.

Dealing with confounding by exclusion, i.e. examining the possibility of carcinogenicity of a more minor type among cancer specimens that do not contain a more important type, becomes a problem of misclassification. This main epidemiologic criterion used for classification of an HPV type as a carcinogen, finding the HPV genotype as a single infection in a cervical scrape or biopsy specimen in a woman with cancer, might sometimes be too lax and prone to error. Colposcopic biopsies and cytology specimens can be misdirected and fail to obtain the critical cells, while contamination of scrapes and biopsies from lower-grade lesions that often surround cancers can detect types other than the causal one. Studies relying on testing of microdissected cervical malignancies will address these issues, but large-scale highly accurate data are not yet available.

Difficulty with control selection adds another level of complexity in assessing carcinogenicity. Cervical cancer typically follows age infection by decades. HPV transmitted at young ages usually become undetectable by DNA or RNA assays and no sensitive serologic assay exists to measure HPV exposure. Consequently, odds ratios based on a comparison of HPV prevalence at the time of case diagnosis to age-matched HPV point prevalence in controls do not estimate true relative risks.

There is not much type-specific prospective data on the carcinogenicity of individual HPV genotypes. The available studies have categorically shown the unique carcinogenicity of HPV 16 and, to a lesser extent, HPV 18 [9, 12]. Khan et al. observed a risk for the remaining women positive by hc2 after excluding those positive for HPV 16 or HPV 18 (including an unknown mix of the types 31, 33, 35, 39, 45, 51, 52, 56, 58, 59 and 68) of only 3.0% (1.9–4.2) compared with 0.8% (0.6–1.1) among women who were HPV negative at baseline. Thus, there is not strong and convincing long-term prospective evidence for individual HPV types other than HPV16 and HPV18.

Finally, the accuracy of detection of HPV genotypes differs between the major PCR-based systems used to generate most of the data [13]. Each of the systems shows differential sensitivity, and some have exhibited cross-reactivity of detection. These detection issues are not critical for evaluating the most important HPV types, but make it difficult to clarify the role of the most weakly carcinogenic and least common ones.

With these caveats, the cervical carcinogenicity of the HPV types listed above varies in strength in a continuum without clear breakpoint, from extremely strong (i.e. HPV 16 and, to a lesser degree, HPV 18) to weak, but still probably carcinogenic in rare instances (e.g. HPV 68, see below). Evaluators taking one extreme position could claim that there is reasonable evidence for carcinogenicity of virtually all the types in the species listed above, extending further the list established in Monograph 90. Strict interpreters of causal criteria could argue for a return to a much more limited list. But based on current evidence, no cut-point between sufficient, limited, and inadequate epidemiologic evidence is entirely defensible.

When epidemiology serves as the science informing public health policy, its limitations must be acknowledged. Weak causal associations with one HPV type are extremely hard to prove in the presence of powerful confounding by strong carcinogens like HPV16. A coming generation of intensive molecular studies of microdissected cancers, with examination of transcriptional activity, may clarify the borderline between carcinogenic and non-carcinogenic HPV types. For public health purposes, the main agents that merit consideration in screening tests and vaccines have already been identified.