Abstract P3-05-02: A minimal DNA-methylation signature to estimate tumor content and molecular subtype in breast cancer tissue samples with potential application to liquid biopsy

Abstract

Abstract Introduction: Breast cancer (BC) is a heterogeneous disease characterized by different molecular "intrinsic" subtypes (PAM50) with distinct prognostic and therapeutic implications. DNA methylation (DNAm) is one of the most studied epigenetic mechanisms involved in BC tumorigenesis and progression. Recently, many studies have demonstrated the capability of DNAm-based assays in detecting and monitoring the disease in liquid biopsy samples from patients with BC, thus confirming promising role of DNAm-based biomarkers in the context of precision oncology. Here we present a novel computational method that exploits a minimal set of informative, clonal DNAm sites to estimate tumor content and molecular subtype in BC tissue samples with potential application for liquid biopsy analyses. Methods: DNAm data of TCGA-BRCA comprising 737 tumor samples, including 135 Basal-like (Basal), 46 HER2-enriched (HER2e) and 556 Luminal A/B (LumA/B) samples, and 96 normal tissue samples were downloaded from GDC Data Portal. To estimate tumor purity, we selected DNAm sites based on the following characteristics: (1) AUC>0.8 or AUC<0.2 (hyper- and hypo-methylated sites, respectively) in tumor versus normal samples; (2) range of beta-values in tumor samples greater than 0.5; (3) max and min beta-values above 0.9 or below 0.1, for hyper- and hypo-methylated sites, respectively; (4) the 3rd and 1st quartiles of normal samples beta-values below 0.3 or above 0.7, for hyper- and hypo-methylated sites, respectively. Tumor purity was then estimated taking the median of beta-values for hyper and 1-beta for hypo of selected sites for each sample. This selection was carried out considering all TCGA-BRCA samples and by PAM50 subtype (Basal, HER2e and LumA/B). After correction for purity, the set of subtype-specific sites (n=89) was also used as input to a generalized multinomial regression model (GLMNET) to classify BC samples into molecular subtypes. The model was trained in the TCGA-BRCA dataset using a 5-fold cross-validation. The performance was then tested in two independent datasets retrieved through the recountmethylation R package: GSE72251 (n=117) and GSE84207 (n=254). Results: Our estimates of sample purities, exploiting only 20 DNAm sites, show high correlation with those from InfiniumPurify (R=0.90, p < 10-16), another DNAm-based tool that requires hundreds of sites for purity estimation. The correlation was further improved using subtype-specific sites (LumA/B: R=0.95, p < 10-16; HER2e: R=0.96, p < 10-16; Basal: R=0.93, p < 10-16). For classification of tumor samples into molecular subtypes, our model obtained high accuracy for all subtypes (99%, 84% and 99%, for Basal, HER2e and LumA/B, respectively) in the training dataset. In the test datasets the F1-scores for Basal, HER2e and LumA/B subtypes were 89%, 60% and 87% for GSE72251, 91%, 42% and 93% for GSE84207. Conclusions: Our computational framework exploiting a minimal signature of DNAm sites demonstrates high accuracy in quantifying tumor purity and predicting molecular subtypes in the training dataset. In the two independent test sets including >350 samples, high accuracy was obtained for LumA/B and Basal samples while HER2e samples showed lower F1-score due to a drop of recall, though obtaining precision >80% in both datasets. Based on these results, a DNAm-based targeted-seq assay has been designed and will be applied to a series of plasma samples from patients with metastatic BC collected within our center. First results of the assay will be presented during the meeting. Citation Format: Dario Romagnoli, Francesca Galardi, Francesca De Luca, Chiara Biagioni, Erica Moretti, Laura Biganzoli, Ilenia Migliaccio, Luca Malorni, Matteo Benelli. A minimal DNA-methylation signature to estimate tumor content and molecular subtype in breast cancer tissue samples with potential application to liquid biopsy [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P3-05-02.