868P - A DNA methylation classifier to predict recurrence from clear surgical margins

Background

Many patients with oral (OSCC) and oropharyngeal squamous cell carcinoma (OPSCC) experience disease recurrence after surgery with curative intent despite R0 resection and adjuvant treatment. Several methods, including frozen section analysis, are available to evaluate surgical margins. However, precancerous epigenetic disturbances may predate morphological changes. This study aimed to address the unmet need for intraoperative recurrence prediction based on molecular features in clear resection margins.

Methods

Infinium EPIC BeadChip 850k methylation analysis of 71 OPSCC primary tumors (TU), 16 contralateral healthy mucosa (HM), and 70 resection margins (RM: OPSCC & OSCC) was conducted. Oncogenic features were selected based on TU vs. HM differential methylation analysis (Kruskal Wallis Tests: η² > 0.14 and FDR < 0.05). Baseline classifiers were trained, and the best-performing classifier was optimized by hyperparameter tuning and feature selection, allowing the integration of 46 TCGA R0 margin (normal adjacent tissue) data. Features selected were checked for common single nucleotide polymorphisms (SNP), and correlation analysis between loci methylation and gene expression (RNAseq) was conducted. Four margin samples analyzed with Epic BeadChip were also sequenced with Oxford Nanopore Technology (ONT) Mk1C.

Results

Forty-nine thousand features were selected to train six baseline classifiers, and XGBoost was chosen. XGBoost training with 10-fold cross-validation, hyperparameter tuning, and seven features (hypomethylated in TU compared to HM) selection led to the construction of a classifier model that predicts recurrence with a mean area under the ROC curve (AUC) of 0.80 (95% CI = 0.73 - 0.87). None of the seven loci used in constructing the model harbored a common SNP (MAF < 0.01). A significant correlation between the methylation level of five of the CpG loci and RNA expression was discovered (p < 0.05). ONT methylation output exhibited a strong positive correlation with Epic BeadChip for all the intersected loci (r = 0.83, p < 0.001), but most importantly, the seven selected features (r = 0.90, p < 0.001).

Conclusions

Disease recurrence can be predicted from morphologically clear surgical margins using a methylation classifier.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Research Training Group GRK-2254 (HEIST, 288342734) funded by Deutsche Forschungsgemeinschaft (DFG).

Disclosure

S. Laban: Other, Institutional, Advisory Board: Merck Sharp & Dohme, Bristol Myers Squibb, Sanofi Genzyme; Financial Interests, Institutional, Advisory Board, Travel reimbursement: AstraZeneca ; Financial Interests, Institutional, Other, Travel reimbursement: Merck Serono. All other authors have declared no conflicts of interest.