Accurate and Efficient Detection of Nasopharyngeal Carcinoma Using Multi-Dimensional Features of Plasma Cell-Free DNA

Abstract

Background: The incidence of Nasopharyngeal carcinoma (NPC) is rising in recent years, especially in some non-developed parts of the world. Hence, cost-efficient means for sensitive detection of NPC are vital.

Methods: We recruited 646 participants, including healthy individuals, patients with benign nasopharyngeal diseases, and NPC patients for plasma cell-free DNA(cfDNA), which underwent low-depth whole-genome sequencing (WGS) to extract multi-dimensional molecular features, including fragmentation pattern, end motif, copy number variation(CNV), and transcription factors(TF). Based on these features, we employed a machine learning algorithm to build prediction models for NPC detection.

Results: We achieved a sensitivity of 95.8% and a specificity of 99.4% to discriminate NPC patients from healthy individuals.

Conclusions: This study can be a proof-of-concept for these multi-dimensional molecular features to be implemented as a noninvasive approach for the detection and even early detection of NPC.

Keywords: cell‐free DNA; copy number variations (CNV); machine learning; nasopharyngeal carcinoma; whole genome sequencing (WGS).

FIGURE 1

General workflow of this study.

FIGURE 2

Fragmentation profiles of cfDNA samples from different cohorts. Fragmentation profiles of cfDNA samples from healthy cohort, nasopharyngeal benign disease cohort, and NPC cohort. The X‐axis represents 5 Mb bins across the human genome, while the Y‐axis represents the difference to the average ratio of short to long cfDNA fragments for each bin.

FIGURE 3

Motif feature of three cohorts on the top 10 selected motifs by proportion. The abscissa represents the motif sequence, and the ordinate represents the proportion of each of the 10 motifs. [Color figure can be viewed at wileyonlinelibrary.com]

FIGURE 4

CNV profiles of cfDNA samples from different groups. The X‐axis represents 0.5 Mb bins across the human genome, while the Y‐axis represents the CNV log2 ratio values in each bin.

FIGURE 5

TFs analysis. The read depth of TFBSs in the healthy cohort (curves in green), nasopharyngeal benign disease cohort (curves in yellow), and NPC cohort (curves in red). [Color figure can be viewed at wileyonlinelibrary.com]

FIGURE 6

Machine learning models detect NPC patients with high sensitivity and specificity based on cfDNA fragmentation profiles. Green ROC curve represents the prediction model to discriminate NPC patients from healthy individuals. Blue ROC curve represents the prediction model to discriminate NPC patients from non‐cancer individuals. Yellow ROC curve represents the prediction model to discriminate NPC patients from benign nasopharyngeal diseases. [Color figure can be viewed at wileyonlinelibrary.com]