1251P - Development of a deep learning algorithm for lung cancer diagnosis using methylation and fragment size profiles of cfDNA

Background

Abnormal methylation of cell-free DNA (cfDNA) enables the diagnosis of lung cancer. In this study, we aimed to identify the best methylation marker for distinguishing lung cancer from healthy individuals, and then develop a deep learning algorithm using methylation and fragment size profiles (MFS) of cfDNA in that region.

Methods

We obtained 828 tissue samples from lung cancer patients, 74 adjacent normal tissue samples, and 656 samples of normal whole blood from the TCGA database and generated genome-wide tissue and plasma methylome data from healthy individuals and lung cancer patients using 215 Methylated DNA immunoprecipitation sequencing (MeDIP-Seq) and 24 Enzymatic Methyl-seq (EM-seq). A total of 366 DNA methylation markers were selected based on the methylation profiling of these datasets, and a targeted EM-seq panel was designed on these markers. The panel was used to produce 142 lung cancer samples (stage I: 26%, II: 11%, III: 42%, IV: 11%, unknown: 11%) and 56 healthy samples. MFS was trained on hyper- and hypo-methylated regions using convolutional neural networks. To confirm the limit of detection (LOD), sequencing was performed on diluted samples of healthy and lung cancer cfDNA with different ratios (tumor fractions of 0.1%, 0.5%, and 1%).

Results

We found that the performance was better in hypo-methylated regions than in hyper-methylated regions. The lung cancer detection performance in the test and external sets reached an accuracy of 81.5% (CI: 78.2% to 87.0%) and an AUC of 0.87 (CI: 0.83 to 0.94). At a specificity of 80%, the sensitivity for cancer detection was 77.5% (CI: 67.3% to 99.4%). The sensitivity for Lung Adenocarcinoma and Lung Squamous Cell Carcinoma at 80% specificity was 61.1% (CI: 42.7% to 94.4%) and 85.7% (CI: 71.4% to 100.0%), respectively. With serial dilution results, we detected down to 0.1% tumor fraction with a specificity of 80%.

Conclusions

This study developed and validated a methylation marker and deep learning model that can distinguish between lung cancer patients and healthy individuals using methylation and size information. This approach provides a sensitive and accurate diagnostic tool for lung cancer.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

GC Genome.

Funding

Ministry of Health and Welfare.

Disclosure

All authors have declared no conflicts of interest.