26P - Multi-feature cell free DNA analysis and ensemble machine learning for early detection of cancer

Background

Early cancer detection is crucial for reducing mortality rates, but current screening methods vary widely in age, intervals, and invasiveness. Unfortunately, over 50% of cancer deaths occur in types without recommended screening tests. Non-invasive multi-cancer early detection (MCED) technology could provide a solution.

Methods

We enrolled both healthy individuals (398) and patients diagnosed with stage I-IV colon (107), liver (113), lung (213), prostate cancer (92), gastric (100), pancreatic (113), breast (74), and ovarian (87) cancers for the development of the Multi-Cancer Early Detection (MCED) test. Whole-genome methylation sequencing was performed on tumor and normal tissues at 15X coverage for marker selection, while cell-free DNA was sequenced at 30X coverage to construct the machine learning model. We calculated three genome-wide features: methylation, copy number variation, and fragment-based patterns. For the training set, which comprised 60% of the samples, support vector machine (SVM) algorithms were applied to these features, and ensemble logistic regression was employed to identify cancer signals and tissue-of-origin (TOO) based on scores from the single-feature models. To determine the minimum yield that maintains equivalent performance, we downsampled a 50% fraction of reads from each cfDNA sequencing dataset.

Results

The overall sensitivity of the cancer detection model was 85.7% at the specificity of 95.6%, and TOO accuracy was 81.1%. The sensitivity performance for cancers without recommended screening tests previously, such as pancreatic (83.9%) and ovarian (79.7%) cancer, has also been maintained at a high sensitivity level. In-silico analysis confirmed that reducing coverage to 15X, half of the original, maintains high performance in sensitivity and specificity, significantly lowering data processing requirements.

Conclusions

Proposed MCED method for eight types of cancer, including pancreatic, and ovarian cancer, which were previously difficult to diagnose early, performs with high sensitivity. The reduced coverage of 15X can lower sequencing costs and increase patient accessibility.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

The National Research Foundation, Republic of Korea.

Disclosure

All authors have declared no conflicts of interest.