Assessment of six markers for cervical (pre)cancer detection 47 2 DNA ISOLATION, MODIFICATION & QMSP DNA isolation, sodium bisulphite treatment and multiplex qMSPs were performed as described previously for markers GHSR, SST and ZIC1 34 and ASCL1, LHX8 and ST6GALNAC5 35. Multiplex qMSPs were designed as described by Snellenberg et al. 38. This assay type is able to detect small amounts of methylated DNA in a background of unmethylated DNA, targets simultaneously multiple genes and provides high sample throughput 38. To verify DNA quality and successful bisulphite conversion, the housekeeping gene β-Actin (ACTB) was used as a reference gene in all qMSPs. Methylation levels were normalised to ACTB using the quantification cycle (Cq) values (2-ΔCq ×100) to obtain ΔCq ratios 39. Part of the methylation data were derived from previous studies: data from 88 cervical samples, including 42 cancers and 46 CIN3, were reported before for methylation markers GHSR, SST and ZIC1 26, 40. DATA & STATISTICAL ANALYSIS Differences in DNA methylation levels between the disease categories were assessed using the Kruskal-Wallis omnibus test followed by post hoc pairwise Wilcoxon-MannWhitney U-tests with Bonferroni adjustment for multiple testing. Bivariate associations between markers were evaluated by Spearman correlation coefficients and presence of multicollinearity in the data was assessed using the variance inflation factor. To compare the individual discriminative performance of each marker in controls and CIN3, we performed simple (univariable) logistic regression for each marker separately. In order to evaluate the improvement in terms of prediction performance of a marker panel involving two or more markers, taking into account the multicollinearity in the data, we performed least absolute shrinkage and selection operator (LASSO) logistic regression. LASSO logistic regression has an inbuilt parameter (penalty parameter) that controls the number of variables eventually included in the model, balanced against model fit. The risk of CIN3 was calculated for each sample as a predicted probability (value ranging from 0 to 1). The predicted probabilities obtained were visualised using the receiver operating characteristics (ROC) curve. The model performance was evaluated by the AUC and corresponding sensitivity and specificity at the best threshold, which is the threshold that maximises the sum of sensitivity and specificity (i.e., Youden J-index). Sensitivities at a threshold corresponding to a predefined specificity of 70% and 80% were calculated for all models. All models were trained to discriminate between controls and CIN3 and subsequently evaluated by leave-one-out cross-validation (LOOCV). Cancer samples were not included in the training; but once trained, the models were tested on these samples as well. All statistical analyses and visualisations were performed using square-
RkJQdWJsaXNoZXIy MjY0ODMw