The primary cause of cervical cancer is persistent infection with human papillomaviruses (HPV) (1–3). Yet, despite the availability of effective, prophylactic vaccines against HPV (4, 5), cervical cancer remains the fourth most common cancer and fourth leading cause of cancer death among women worldwide (6). In low Human Development Index settings, it ranks second for both measures (6). This is due in part to the fact that current vaccines cover only nine of the most oncogenic HPV types out of over 200 types that have been identified (3). More importantly, however, availability of the vaccine is relatively recent (2006) and is only approved for individuals below the age of 45. Additionally, even within the guideline-approved age groups, vaccination rates remain low. Together, these facts mean that cervical cancer prevalence and mortality will not decrease significantly in the short term (7–9). Thus, general population screening for cervical cancer must continue for the foreseeable future.

General population cervical screening relies on estimating the risk of high-grade lesions or cancer from tests for cervical cell infection by high-risk HPV (hrHPV) types and cervical cytology (Papanicolaou or Pap) tests (10–12). Risk-based management then dictates whether a woman proceeds under general screening, more frequent surveillance screening, or referral for colposcopy, histologic diagnosis, or treatment (11, 12). While these processes have reduced cervical cancer incidence and mortality through the early detection and treatment of cervical precancerous lesions and cancer (10, 13–15), testing improvements providing refined risk estimation would offer significant benefits. The cytology test has low sensitivity to detect precursor lesions, while a single hrHPV test cannot distinguish between a persistent infection and one that the immune system will naturally clear (16). Furthermore, among women who are hrHPV-positive and/or have abnormal cytology, testing cannot identify those who will progress to high-grade cervical lesions or cancer if not treated (17–22). Thus, reliable cervical cancer prevention requires frequent re-testing, referral for colposcopy and possible histologic diagnosis for women in higher risk intervals, and a recommendation for excisional treatment for women with a risk of cervical intraepithelial neoplasia grade 3 (CIN3) of 26% or higher (12). A consequence of this protocol is that women in under-resourced settings or hard-to-reach communities (e.g., immigrant or rural communities) are often under-screened and therefore at risk of undetected precancerous lesions and cancer (23–26). In contrast, in adequately screened populations, there is a potential for overtreatment, with its associated risks and complications (27–31). Methods to improve risk assessment for cervical cancer, particularly assays that can predict regression of precursor lesions or clearance of HPV infection, would be expected to benefit both populations. Improving the effectiveness of less-frequent testing could enable improved strategies for delivering screening to under-screened populations, thereby reducing cervical cancer incidence and mortality (23–26). In screened populations, improved risk assessment could reduce overtreatment.

We hypothesized that profiling adaptive immune responses, particularly T cell responses, could provide an additional axis along which to assess CIN3 and cancer risk due to the role of these responses in clearing HPV infections (32, 33) and resolving both low- and high-grade lesions (34–38). Indeed, such T cell responses are also associated with positive outcome even after cervical cancer has developed (37, 39–43). In order to profile these responses, we selected high-throughput adaptive immune receptor repertoire sequencing, because it has become one of the primary techniques for profiling adaptive immune responses in cancer (44) and for identifying potential immune biomarkers (45–50).

The goal of this study was two-fold. First, we aimed to determine whether T cell receptor (TCR) repertoire profiling could be conducted on cervical cytology samples, the sample type routinely collected for cervical cancer screening and surveillance and on which cytology tests and hrHPV tests are conducted. The ability to conduct TCR repertoire profiling on this sample type would greatly facilitate incorporation of TCR profiling into current screening protocols. Second, we aimed to determine whether predictive signatures could be identified in these repertoires, signatures that are predictive of a patient’s cytology/hrHPV test results at their next round of testing. We are aware of only two papers that have described TCR repertoire profiling in the context of cervical lesions or cancer, and both differ from the current study in important ways. Lang Kuhs et al. (51) compared peripheral blood TCR repertoires between women who had cleared an HPV-16 infection and women who had HPV-16-related CIN3 or higher (including invasive cancer cases). Cui et al. (52) compared the peripheral blood repertoires between three groups of women: women with cervical cancer, women with CIN of any grade, and healthy women. Cui et al. (52) additionally sequenced tumor and lymph node repertoires from women with cervical cancer. Both studies compared repertoire-level summary statistics, such as diversity and gene segment usage. Cui et al. (52) additionally examined repertoire overlap between samples and between groups. While TCR repertoire sequencing from blood, tumor, and lymph node samples is routine, we are the first to demonstrate successful TCR sequencing from cervical cytology samples. This is important not just because such samples are already routinely collected for cervical screening and surveillance, but also because the cervical repertoire is expected to be more representative of the T cells acting in the local immune response against HPV infection and abnormal lesion cells (53). Furthermore, while both studies (51, 52) make the important contribution of demonstrating differences in peripheral blood T cell repertoires between stages along the cervical cancer spectrum, these differences are reflective of different phenotypes possessed by the donors at the time of sample collection and are not predictive of the patients’ disease trajectories.

In this proof-of-principle study, we demonstrate the ability to obtain sufficient DNA for TCR repertoire profiling from cervical cytology samples collected during general and surveillance screening. Additionally, a subset of the women have had a second cervical sample collected 280-711 days after the one on which we conducted sequencing ( Figure 1 ), allowing us to determine whether they experienced a transition to cytology/hrHPV test results expected to have a lower CIN3 or cancer risk compared to the index results. Indeed, we found that women with lower risk at follow up had evidence of more robust T cell clonal expansion in the index sample than women with persistent risk. Furthermore, applying a machine learning-based method we previously developed (45–47, 50), we identified a biophysicochemical motif in the complementarity-determining region 3 (CDR3) of TCR β chains (TCRB) whose presence was associated with lower-risk follow-up test results with 95% accuracy by leave-one-out cross-validation (19 of 20 trajectories correctly predicted). These results suggest that sequencing TCR repertoires from the cervical samples used for cytology and hrHPV testing could improve cervical cancer screening by further refining the risk estimates used for personalized risk-based management. Furthermore, these results represent a significant advance in the field of immune repertoire predictive modeling as the first to demonstrate the potential to predict a clinically relevant future outcome.

The need to improve triage of abnormal cytology and hrHPV test results during cervical cancer screening and surveillance motivated us to undertake this proof-of-principle study to determine whether T cells present in the cervical samples used for those tests could be used to refine estimates of cancer progression risk. The first goal was to demonstrate that TCR repertoires could be characterized from these samples and to provide the first such characterization. To this end, we have shown that we can obtain robust TCR repertoire data from cervical samples ( Supplementary Figure 1 ). We have further shown that the commonly used repertoire-level summary statistics do not differ between the five result categories and do not differ between HPV+ and HPV- samples ( Supplementary Figures 1 – 4 ). Thus, while we observed quite a bit of sample-to-sample variability in these measures, we find no evidence that there are differences between the result categories or HPV infection states in terms of lymphocyte presence or clonal expansion. We conclude that samples from all result categories are expected to be equally good candidates for the approach we are proposing.

Our second goal was to determine whether TCR repertoire or sequence properties were associated with the cytology and hrHPV test results obtained on the patient’s first follow-up sample. Thus, for the 24 of 103 women who had a second cervical sample collected following the one from which we sampled T cells, we grouped them according to whether or not the follow-up test results were of lower risk than those of the index results ( Table 1 ). We found that, among the widely used repertoire-level summary statistics, there were no differences between women who transitioned to lower risk versus those of persistent risk except for a difference in the median largest clone size ( Figure 5 ). We found that for women who transitioned to lower risk results, the size of the largest clone in the sample was much larger on average than for women who did not, suggesting that their T cell repertoires had undergone more extensive clonal expansion for at least one clone in the local cervical T cell response. While finding a difference in the median largest clone size may seem to contradict the lack of a difference in median clonality, we suspect this is due to the fact that clonality is a repertoire-level measure that attempts to quantify the extent to which the distribution of all clonal relative abundances in a repertoire departs from a uniform distribution. That is a different question than asking whether the median size of the largest clone is different between two groups of repertoires. We did, however, find that, while the median clonality was not statistically significantly different between the two groups, there is a trend toward higher clonality for the group of women who transitioned to lower risk results ( Figure 5F ). Both the larger median largest clone size and the trend towards higher clonality could be indicative of a more robust or targeted antigen-driven immune response. Despite this difference, however, there is no largest clone size that completely separates the two groups of women, as the lower range of values for women who transitioned to lower risk results overlaps the full range of values for women who did not. This is analogous to the result by Cui et al. (52) who found that the average number of unique amino acid TCR sequences in sentinel lymph nodes differed between women with cervical cancer who experienced disease progression versus though who did not, but despite this difference, there was significant overlap in the ranges of values between the two groups of women making it difficult to gain any predictive capability from the measure.

Based on the above results, we concluded that, while a high largest clone size indicates a high likelihood of transitioning to a lower risk combination of cytology and hrHPV test results, a low largest clone size appears to be equally associated with both lower and persistent risk at follow up. Thus, to increase the accuracy of identifying women likely to persist, we built on our prior work developing statistical classifiers that accurately classify TCR repertoires according to sequence-level properties present within the repertoire (45–47, 50). The basis of our method is identification of a biophysicochemical motif in TCR CDR3 sequences that is shared across samples with a common phenotype. In this study, we identified a motif whose presence in TCRB CDR3 sequences marked cervical repertoires collected from women who later transitioned to lower risk cytology/hrHPV result categories. Similarly, the absence of this motif marked cervical repertoires collected from women whose risk level persisted. The motif served as a biomarker of future lower risk with 96% accuracy (23/24) by exhaustive leave-one-out cross-validation in which all sequences from a sample are held out each round ( Figure 6A ). The accuracy calculated over the 20 women who did not receive treatment ( Table 1 ) is 95% (19/20). The average classification accuracy over 20 permutation runs was 62%, and all runs had an accuracy < 95% ( Supplementary Table 2 ). Thus, we conclude that the probability of observing a cross-validation classification accuracy of 95% or higher by chance is p < 0.05.

Both the 96% and the 95% accuracies are a result of misclassifying a single sample, 3_40. This sample was misclassified according to the labels we assigned and used for training the model. We classified this woman as having persistent or increased risk, because the index cytology result was ASCUS and the follow-up cytology result was LSIL ( Table 1 ). In contrast, the model assigned a probability of 0.99 that this woman would experience a transition to lower risk, and indeed she transitioned from hrHPV+ to hrHPV- ( Table 1 ). Thus, she may indeed be transitioning to lower risk cytology/hrHPV result combinations.

In the model, we represented amino acid 4-mers from TCRB CDR3 sequences according to their biophysicochemical properties using Atchley factors in order to capture properties of the CDR3 sequence that determine the CDR3’s peptide-binding specificity, as motivated in (46). The shared motif is described by the model weights indicating the relative importance of the different amino acid properties, different 4-mer positions, and the 4-mer’s abundance in the sample to the calculated probability that the 4-mer is derived from a TCRB sequence found in a repertoire sampled from a woman who transitioned to lower risk ( Figure 6B ). We found that 4-mers composed of four small, hydrophilic residues result in an increased probability of a reduced risk classification. The probability of a reduced risk classification is further increased for 4-mers composed of residues that tend to form alpha-helices in the first and second 4-mer positions and positively charged residues that tend to form bends, coils, or turns in the third and fourth positions. Finally, we found that higher 4-mer relative abundance in a sample also increased the probability of a reduced risk classification. The properties of this motif are distinct from the three other TCR motifs we have identified with this method, and indeed all four motifs are distinct (46, 50).

We hypothesize that the motifs we have identified mark TCRBs that have undergone clonal expansion as part of an antigen-specific immune response critical to the resolution of cervical lesions and/or clearance of hrHPV infection. This hypothesis is supported by the finding that in all women with high-scoring motifs, the clone bearing the woman’s top-scoring 4-mer/motif is one of the most abundant clones in the sample ( Supplementary Table 5 ). Indeed, in many cases, it is the most abundant. The most obvious hypothesis is that the relevant antigen is an HPV antigen and the corresponding immune response critical to clearing infection. An intriguing alternative is that the motif is associated with the presence of particular microbes in the cervical microbiome, such as Lactobacillus iners, which has been associated with the clearance of hrHPV, or Gardnerella, which has been associated with progression (66). Finally, there could be antigens expressed on cells with abnormal cytology and the corresponding immune response critical to eliminating these abnormal cells. A comparison of the motif-bearing CDR3 sequences ( Figure 6C ) against databases of curated TCR sequences with known specificities revealed no matches (67–69). Thus, experimental studies will be required to determine whether these motifs mark a TCR specificity relevant to the process of lesion resolution or infection clearance, and, if so, what that antigen is.

While the motifs we have identified are associated with a transition to lower risk, both Lang Kuhs et al. (51) and Cui et al. (52) identified sequences associated with CIN or cancer. Lang Kuhs et al. observed that the TCRBV6-7 gene segment was overrepresented among women with HPV16-associated CIN3 (51). Thus, we compared the TCRBV6-7 relative abundance between cervical cytology samples with a normal versus HSIL reference diagnosis, between women who transitioned to lower risk and women whose risk level remained unchanged, and between women who were HPV16+ versus HPV16-. In all cases, we found no difference. We further found that none of our motif bearing CDR3s ( Supplementary Table 4 ) were part of rearrangements that included TCRBV6-7. Key differences between our study and the Lang Kuhs et al. study that could explain the different results is that our study included women infected with a variety of HPV types and our repertoires were derived from cytology rather than blood samples. Cui et al. (52) found CDR3 sequences that were shared exclusively among patients with CIN or exclusively among patients with cervical cancer, with sharing ranging from 9-11 and 8-9 patients, respectively. None of these CDR3 sequences were found in any of our sampled repertoires. This is not necessarily surprising. Cui et al. had sample sizes of 30 CIN patients and 25 cancer patients. Thus, even the most broadly shared CDR3 sequences were not observed in more than half of the patients. Furthermore, our failure to observe these sequences may be a consequence of sample the cervical rather than peripheral blood repertoire.

There are multiple factors with the potential to confound our results, differences between the two groups of women with respect to race/ethnicity, age, HPV infection, time between the two sets of results used to determine disease trajectory, and whether treatment was received between them. We have controlled for or assessed the potential impact of each of these as much as possible. First, our patient cohort consists entirely of White Hispanic women. This not only controls for the known confounding factor of race/ethnicity in cervical cancer risk but also provides a valuable data resource for future studies aimed at elucidating the causes of cervical cancer disparities. Regarding age, while we find that women with an LSIL diagnosis for the reference cytology result have a significantly lower median age than women with a normal reference cytology result ( Figure 3A ), we found no difference in the median age between women who transitioned to lower risk on follow-up and those who did not ( Figure 4B ). We found no difference between the two groups of women for the amount of time between the first and follow-up cervical cytology/hrHPV results, or the number of HPV types with which an individual was infected ( Table 1 , Figure 4 ). We further found no difference in HPV16 or HPV18 positivity between the two groups ( Table 1 ). Finally, receipt of excisional treatment for precancer also does not appear to explain the difference, given that only four of 24 patients received treatment between the two sets of results ( Table 1 ). Among the 20 untreated women, there are 14 in the reduced risk category and 6 in the persistent risk category, and the leave-one-out cross-validation accuracy excluding the four treated patients is 95% (19/20). Overall, our sample, while small, has similar representation for each of these factors in both groups, and/or the groups themselves do not differ on the basis of these factors. Thus, we conclude that other factors are contributing to the difference in apparent disease trajectory between the two groups sampled in this study. Additional potential confounders that we were unable to account for include HLA type, smoking status, and a history of HPV vaccination. These need to be accounted for in a future study with a larger sample size.

All four of the women who received treatment between their reference and follow-up cytology/hrHPV results received lower risk results on follow-up. They were also predicted to have lower follow-up risk by our model. It is important to note, however, that their transition to lower risk may be driven solely by the treatment and may not reflect the biological state of their lesion, HPV infection, or T cell response at the time of sample collection. Only ~9% of women receiving excisional treatment show residual or recurrent abnormal cytology or persistent hrHPV infection after treatment, compared to 40% to 80% of untreated women, depending on age and lesion grade (70).

The above-presented evidence from the 20 untreated women is consistent with our original hypothesis that women who experience spontaneous regression of cervical lesions and clearance of HPV infection do so, at least in part, as a consequence of an antigen-specific, T cell-mediated immune response that can be detected via TCR sequence patterns. There are limitations of our study, however, the primary ones being the small training set size and the lack of a test set. We fully expect the result here to hold, however, given that our prior results with similarly sized training sets have shown model generalizability on test data (45, 50). Indeed, the T cell study (50) had 95% accuracy by cross-validation and 80% accuracy on an independent test set. More importantly, we note that there was no model selection in the current study. Here, we used the model selected in (46) and only used the training set to fit the model parameters ( Figure 2 ). Thus, we expect test set accuracy greater than 80%. Additionally, we expect much higher accuracy with a large training set (71). Indeed, the purpose of this proof-of-principle study was precisely to determine whether a large-scale study is justified, and the results presented here do indeed warrant follow up in a large study. Unfortunately, given the longitudinal nature of our study design and the relative rarity of disease trajectories of persistent risk, such a study is expected to take many years.

One potential limitation of our study is the use of cytology results without histologic confirmation. We used cytology rather than histology results, because, per guidelines, histologic diagnosis is only recommended for a subset of abnormal results. Thus, in this proof-of-principle study using standard-of-care data, cytology results (plus reflex hrHPV test results) are the only result types consistently available for all women across all abnormal and normal result combinations. Indeed, there would be ethical questions around designing a research study in which histologic confirmation was obtained for all results, even just for all abnormal results. We do, however, provide histologic diagnoses as well as repeat follow-up cytology and hrHPV test results where those are available ( Supplementary Table 6 ). Furthermore, risk-based management relies heavily on cytology and hrHPV test results for clinical decision making. Thus, our goal is to maximize the amount of information about risk that can be obtained from cervical cytology samples.

Additional limitations of our study include the use of cytology samples that had been stored at room temperature in PreservCyt solution for five weeks before processing, the focus on a single race/ethnicity, and the focus on only a single TCR chain (TCRB). We recommend that future studies building on the results presented here utilize freshly collected research samples that would be expected to have better DNA yield as well as to permit RNA-based studies. RNA-based studies could utilize single-cell RNA sequencing which would enable joint analysis of a T cell’s gene expression profile (including T cell type and subset) with its full TCR (both α and β chains). Additionally, replication of this finding across multiple races and ethnicities is needed. The results of this proof-of-principle study provide justification for pursing these aims.

Determination of whether a similar signal can be detected in blood would inform whether development of a blood-based assay would be possible. In our view, given the benefits of conducting all tests on a single sample, pursuing application of this approach to blood would only make sense if it was determined that cervical cytology sampling was no longer needed. We would not expect a blood-only approach to be fruitful given the increased repertoire diversity in blood and the consequent rarity of cervix-associated T cells in blood. This is supported by current evidence indicating that HPV16 reactivity among peripheral blood T cells is not associated with regression in women with HPV16 CIN 2/3 (53). The Trimble et al. (53) study should not be taken as definitive, however, due to the limited follow-up period and the fact that only HPV16 E6/E7 antigens were tested. Other antigens may be relevant, as discussed above. Furthermore, after discovering a signal in the cervical repertoire, optimized approaches for detecting it in blood can be designed. Thus, a study to determine the amount of overlap between the cervical and blood T cell repertoires may be warranted.

The results presented here are significant both because they represent a significant advance in the field of immune repertoire predictive modeling as the first to demonstrate the potential to predict a future outcome, but also because they have the potential to stimulate increased interest in coupling the current focus on lesion formation (cytology testing) and persistent hrHPV infection (hrHPV testing) with factors mediating regression, i.e., a T cell-mediated antigen-specific immune response. Given the high proportion of cases that undergo regression, including an assay that triages samples based on the potential for future regression rather than focusing solely on the current presence of a lesion or its causal agent could significantly improve screening effectiveness with significant benefits both for populations that currently receive guideline-recommended screening as well as for under-screened populations. In screened populations (i.e., in high resource settings), testing is recommended every three to five years unless an abnormal result is obtained, at which point there is a shift to more frequent surveillance testing and a recommendation for treatment of women with a CIN3 risk of 26% or higher (10–12). While this strategy has significantly reduced cancer incidence and mortality in screened populations, it is accompanied by overtreatment in the form of unnecessary procedures that have their own risks and complications, particularly among women of reproductive age. Thus, the ability to reduce the frequency of surveillance screening and the number of unnecessary treatments is expected to reduce this overtreatment burden on screened women (27–31). More importantly, the current source of cervical cancer resides primarily in unscreened or under-screened populations, and here an approach that triages samples based on the potential for future regression has the potential for even greater impact. To deliver effective screening to under-screened populations, there has been a focus on devising strategies for preventing the most cancers via a single screening encounter (23–26). The current strategy focuses on deploying a one-time test at an age where the test is most likely to catch persistent infections and pre-cancers, such as between the ages of 30 and 39 years. This one-time strategy recognizes that it would not prevent cancers occurring before the screening age. The ability to distinguish women likely to regress from those at highest risk of progressing to cancer might allow the one-time test to be deployed at a younger age, thereby preventing a larger number of cancers. Additionally, it could facilitate decision making for screen-positive women allowing focused follow-up on those most at risk of progression. Thus, while the effectiveness of one-time testing remains an open question, the development of assays to help identify women likely to regress is expected improve delivery of screening to under-screened populations by increasing the amount of information about a woman’s risk that is obtained at each screening encounter. Deployment of such an assay would only be feasible if it could be designed for use in low resource settings. However, if this could be accomplished, it would be expected to contribute towards a reduction in cervical cancer incidence and mortality.

The TCRB sequencing data are freely available from the VDJServer Community Data Portal (CDP) (http://vdjserver.org) under the project accession 4995411523885404651-242ac118-0001-012 (56). The study metadata are consistent with the AIRR C MiAIRR Standard (72). VDJServer is part of the AIRR C Data Commons (73).

The studies involving human participants were reviewed and approved by Institutional Review Board of the University of Texas Southwestern Medical Center. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

SC and LC wrote the manuscript. SC, JO, WZ, and BS performed experiments. SC, JT, EL and LC analyzed and interpreted the patient data. LQ handled all IRB document processing and sample retrieval. NM, AG, JT, EL, and LC designed the work. All authors contributed to the article and approved the submitted version.

This research was supported by Simmons Comprehensive Cancer Center Development Funds and by a charitable donation from Young Texans Against Cancer, both to LC and JT.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.