Abstract LB410: Precision oral cancer screening and diagnostic solution using DNA methylation and machine learning to stratify high-risk lesions in saliva from patients of mixed ancestry

Abstract Analysis of quantitative methylation specific PCR (qMSP) data for diagnosis and early detection of cancer has consisted of summarizing singleplex or multiplex DNA data in a cumulative methylation index, followed by threshold analyses. Recently, a novel clustering algorithm was used to examine digital PCR data, prior to downstream analysis. However, clustering of Real-Time qMSP data has not been used by laboratories developing DNA methylation biomarkers for the oral cancer screening and diagnostic space. We created a precision DNA methylation algorithm to quantify Differentially Methylated Promoters (DMPs) with Real-Time PCR instruments, combined with machine learning, for discovery and validation of head and neck squamous cell carcinoma (HNSCC) early detection, diagnosis, and prognostication targets. Analytic validation of PAX1, PAX5, ZIC4, PLCB1, and HHIP was performed to develop a qMSP protocol for clinical samples. The performance of the six singleplex reactions was tested in 307 oral cancer tissue and 55 normal uvulopalatopharyngoplasty (UPPP) samples from a mixed ancestry cohort (40% Black) obtained from the Johns Hopkins School of Medicine Head and Neck Cancer Tumor Bank. An R script for automated analysis of qMSP data was developed to import, process, and analyze multiple qMSP raw data files exported from Applied Biosystems SDS or DA2 software packages. The workflow includes data preprocessing; filtering by quality control metrics, such as CT and PCR efficiency; normalizing against a control gene (Bactin), and visualizing results through boxplots. A precision DNA methylation algorithm was then developed to perform unsupervised hierarchical clustering of singleplex qMSP observations for five genes, center the data, calculate the distance between all samples, determine the variance explained by each Principal Component (PC), set a cutoff DNA methylation value that maximizes performance for each gene and identify the best model fit. Logistic regression, Linear Discriminant Analysis, Loess, K nearest neighbor, and Random Forrest models, as well as an ensemble of all five models were then trained to model the relationship between test samples and PAX1, PAX5, ZIC4, PLCB1, HHIP DMPs. Model performance was compared based on accuracy and logistic regression was used for downstream analyses. The discriminatory ability of the five genes was evaluated using the Receiver Operator Characteristic (ROC) curve and Area Under the Curve (AUC) analyses. The best performance was obtained when using all five genes (PAX1, PAX5, ZIC4, PLCB1, HHIP): 94% Sensitivity, 96% Specificity, 97% Positive Predictive Value, 91% Negative Predictive Value, correctly classifying 95% with an AUC = 0.99. We also found HHIP fully discriminated between normal and tumor samples in a smaller subset of saliva samples (n=73). The normalized tissue-saliva Inter Quartile Range (IQR) ratio of HHIP DNA methylation was 98%. These results warrant to be validated in a larger cohort. RealTime PCR based tests have shown to be cost effective and scalable. The exceptional discriminatory power between normal and cancer taken into the context of post COVID excess installed Real Time PCR instruments, hold a promise of improving oral cancer early detection and diagnostic pipelines worldwide. Citation Format: Ashley Ramos-Lopez, Amanda Garcia Negron, Guie Beeu Guerrero Hunt, Adhi Guerrero-Thillet, Carolina Zambrano Rabanal, Paola Quiñonez Mendez, Andrea Lopez-Marrero, Alvaro Gutierrez, Fernando Zamuner, Bruce J. Trock, Wayne Koch, Mariana Brait, David Sidransky, Rafael Guerrero-Preston. Precision oral cancer screening and diagnostic solution using DNA methylation and machine learning to stratify high-risk lesions in saliva from patients of mixed ancestry [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_2):Abstract nr LB410.