Targeted RNA next generation sequencing analysis of cervical smears can predict the presence of hrHPV-induced cervical lesions

Abstract

Background: Because most cervical cancers are caused by high-risk human papillomaviruses (hrHPVs), cervical cancer prevention programs increasingly employ hrHPV testing as a primary test. The high sensitivity of HPV tests is accompanied by low specificity, resulting in high rates of overdiagnosis and overtreatment. Targeted circular probe-based RNA next generation sequencing (ciRNAseq) allows for the quantitative detection of RNAs of interest with high sequencing depth. Here, we examined the potential of ciRNAseq-testing on cervical scrapes to identify hrHPV-positive women at risk of having or developing high-grade cervical intraepithelial neoplasia (CIN).

Methods: We performed ciRNAseq on 610 cervical scrapes from the Dutch cervical cancer screening program to detect gene expression from 15 hrHPV genotypes and from 429 human genes. Differentially expressed hrHPV- and host genes in scrapes from women with outcome "no CIN" or "CIN2+" were identified and a model was built to distinguish these groups.

Results: Apart from increasing percentages of hrHPV oncogene expression from "no CIN" to high-grade cytology/histology, we identified genes involved in cell cycle regulation, tyrosine kinase signaling pathways, immune suppression, and DNA repair being expressed at significantly higher levels in scrapes with high-grade cytology and histology. Machine learning using random forest on all the expression data resulted in a model that detected 'no CIN' versus CIN2+ in an independent data set with sensitivity and specificity of respectively 85 ± 8% and 72 ± 13%.

Conclusions: CiRNAseq on exfoliated cells in cervical scrapes measures hrHPV-(onco)gene expression and host gene expression in one single assay and in the process identifies HPV genotype. By combining these data and applying machine learning protocols, the risk of CIN can be calculated. Because ciRNAseq can be performed in high-throughput, making it cost-effective, it can be a promising screening technology to stratify women at risk of CIN2+. Further increasing specificity by model improvement in larger cohorts is warranted.

Keywords: Cervical intraepithelial neoplasia; High risk human papilloma virus; Machine learning; Screening; Targeted RNA sequencing.

Fig. 1

Summary of cohorts for analysis. Cohort A was collected at random from HPV-positive tested women, and cytology scores added afterwards; cohort B was selected for similarly sized groups with NILM, LSIL, and HSIL. Cohort C consists of random smears, tested negative for hrHPV-DNA. The table in Fig. 1 relates to cytological outcomes in cohort A

Fig. 2

Positivity rates for hrHPVE6/7 and hrHPV E6*, related to cytology (A) and to colposcopy/histology outcome at follow-up (B)

Fig. 3

Distribution of HPV genotypes over groups of scrapes with different outcome (cytological or histological). Note that a number of HPV genotypes are exclusively found in lower grade lesions in this cohort

Fig. 4

A Unsupervised hierarchal clustering of ciRNAseq data, excluding hrHPV transcript information, of groups defined as safe (NILM at primary and secondary cytology) and CIN2+. The table in B shows significantly overexpressed genes in smears of CIN2+ women, as determined by the Wilcoxon test (all highly significant with adjusted P-values < 0.00002)

Fig. 5

Outcome of the application of a random forest model, generated with ciRNAseq data from 360 smears, on an independent dataset of 63 smears. With a preset cutoff score of 0.8, all samples regarded safe (NILM at first scrape and repeat scrape) were correctly identified