Multimodal analysis of cell-free DNA enhances differentiation of early-stage breast cancer from benign lesions and healthy individuals

Abstract

Background: Breast cancer (BC) remains the second leading cause of cancer-related mortality among women worldwide. Liquid biopsy based on circulating tumor DNA (ctDNA) offers a promising noninvasive approach for early detection; however, differentiating malignant tumors from benign abnormalities remains a significant challenge.

Results: Here, we developed a multimodal approach to analyze cfDNA methylation and fragmentomic patterns in 273 BC patients, 108 individuals with benign breast conditions, and 134 healthy controls. Genome-wide analyses revealed distinct cfDNA copy number alterations and cytosine-enriched cleavage sites in BC patients. Targeted sequencing further revealed unique methylation patterns, including hypermethylation in GPR126, KLF3, and TLR10 and hypomethylation in TOP1 and MAFB. Our machine-learning model achieved an AUC of 0.90, with 93.6% specificity and 62.1-66.3% sensitivity for stage I-II cancers. In symptomatic populations, sensitivities were 50.0%, 68.2%, and 64.7% for BI-RADS categories 3, 4, and 5, respectively, with 96.1% specificity.

Conclusions: These findings underscore the potential of cfDNA biomarkers to enhance BC detection and reduce the rate of unnecessary biopsies.

Keywords: Benign abnormalities; Breast cancer; CfDNA; Methylation and fragmentomic.

Fig. 1

Overview of study design. The plasma samples in the discovery cohort underwent the SPOT-MAS assay. Features of cfDNA (Methylation, End motif, Fragment length, Copy number) were collected from both the target and the genome-wide fraction after sequencing. A multi-feature model was developed using these features to distinguish BC patients from non-cancer (benign and healthy) individuals. To validate this model, we recruited 119 symptomatic participants who had been diagnosed with BI-RADS 3-5 by mammography. The participants in the validation cohort also underwent the SPOT-MAS assay, which was similar to those in the discovery cohort and confirmed by fine needle aspiration (FNA) afterward

Fig. 2

Analysis of genome-wide copy number aberration (CNA) in cfDNA. Scatter plot shows log2 fold change of DNA copy number in each bin across 22 chromosomes of 273 BC patients versus 108 benign patients (A) or 134 healthy subjects (B) in the discovery cohort. Each dot represents a bin identified as gain (red), loss (blue), or no change (grey) in the copy number. Proportions of different CNA bins in each chromosome for BC versus benign patients (C) and BC versus healthy individuals (D). E Venn diagram indicates the number of significant bins overlapped between two pair-wise comparisons (BC versus benign and BC versus healthy). Boxplots showing 2 gain bins (F) and 3 loss bins (G) of copy number in BC patients compared to benign or healthy individuals

Fig. 3

Distinct end motif patterns of plasma cfDNA in BC, benign and healthy individuals. A Heatmap shows log2 fold change of 256 4-mer end motifs in BC patients compared to benign or healthy subjects (end motifs were highlighted in red for higher frequency and blue for lower frequency in BC patients). Boxplots showing 8 increased EMs (B) and 7 decreased EMs (C) of copy number in BC patients compared to benign patients or healthy controls. D Heatmap indicates log2 fold change of motif 21 between BC versus benign patients and BC versus healthy subjects. E Boxplots show 8 significant ME21s in BC patients compared to benign or healthy individuals

Fig. 4

Analysis of targeted methylation and copy number in plasma cfDNA. A Volcano plot shows log2 fold change and significance methylation (−log10 Benjamini- Hochberg adjusted p-value from Mann-Whitney U test) of 450 target regions when comparing BC patients to benign or healthy controls in the discovery cohort. There are 3 DMRs (adjusted p-value < 0.05) and overlapped across these two pairwise comparisons, color-coded by genomic locations (highlighted in red for log2FC>0 and blue for log2FC<0). B Boxplots showing 3 regions (HIVEP2_GPR126, KLF3_TLR10, and MAFB_TOP1) in BC patients, benign patients, and healthy participants. C Heatmap shows log2 fold change of 156 regions with significant copy number in BC patients compared to benign or healthy individuals. The number of significant regions in each pairwise comparison and overlapped regions was indicated via Venn diagram (D). Boxplots showing the top 5 with significantly raised CNA values (E) and the top 5 with significantly reduced CNA values (F) of BC patients compared to benign or healthy individuals

Fig. 5

Model construction and evaluation of multi feature model. A Model construction workflow. Receiver Operating Characteristic (ROC) curves showing the performances of the multi-feature models with the classification BC versus noncancer (B), BC versus benign (C), and BC versus healthy (D). E Bar plot displaying sensitivity and specificity values in the train and the test set

Fig. 6

Performance of the model in BC detection at different stages or with different subtypes. ROC curves showing the performance of the model for BC stage I to III in the train set (A) and the test set (B). C Bar plot showing the sensitivities of the model according to stages I-–III. ROC curves showing the performance of the model on different BC subtypes in the train set (D) and the test set (E). F Bar plot showing the sensitivities of the model according to subtypes of BC patients

Fig. 7

Performance of the model in the validation cohort. A Heatmap shows the demographic details of 119 symptomatic participants in the validation cohort. B Bar plot showing sensitivities and specificities of the model according to BI-RADS lesions. C Bar plot indicates sensitivities of BC patients in different stages. D Bar plot displaying sensitivities of BC patients in different subtypes