1233P - Accurate prediction of gastrointestinal cancer tissue of origin using comprehensive plasma cell-free DNA fragmentomics features

Background

Circulating cell-free DNA (cfDNA) is a promising biomarker for early cancer detection, and its fragmentomics features have been successfully used to detect cancer signals in blood. However, its ability to predict the tissue of origin (TOO) of cancers remains to be evaluated, which is highly desirable to differentiate the most common types of gastrointestinal (GI) cancers, including colorectal (CC), esophageal (EC), gastric (GC), liver (LC), and pancreatic cancer (PC).

Methods

Whole-genome sequencing was performed for the cfDNA of 769 cancer patients (149 CCs, 137 ECs, 149 GCs, 272 LCs, and 62 PCs), to calculate the coverage at repetitive genomic regions (RepeatsCov), the depth and the cleavage diversity around transcription start sites (TSSDepth and TSSClvDiv), and the microbiome abundance (MicrobeAb). Together with other classical fragmentomics features, including copy number variation (CNV), end motif diversity (EDM), fragment size ratio (FSR), and promoter fragmentation entropy (PFE), a stacked ensemble machine learning classifier was trained and tested with sample ratio of 1:1 to predict the TOO of the GI cancers.

Results

The performance of each single feature was evaluated first, showing that the FSR model had the highest accuracy of 67.1% while the RepeatsCov model had the lowest of 53.9%. The ensemble of all the features resulted in an accuracy of 67.6%. Interestingly, a model combining MicrobeAb, RepeatsCov and FSR achieved the highest accuracy of 69.4% for all cancers (CC: 63.8%, EC&GC: 63.3%, LC: 83.6%, and PC: 43.8%), and an elevated accuracy of 87.8% to predict the top two most likely TOOs. We also trained and tested a previously reported multi-features-based model on our data, and our classifier achieved higher accuracy (69.4% vs. 60.6%).

Conclusions

We comprehensively evaluated the classical and our newly developed cfDNA fragmentomics features in predicting the TOO of cancer signals, and showed that by combining features including MicrobeAb, RepeatsCov and FSR, we were able to maximize the accuracy in predicting GI cancers’ TOO. However, results also indicate that features should be carefully selected to avoid multicollinearity or other negative effects.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

National Key Research and Development Program of China.

Disclosure

R. Fu, K. Xie, Y. Liu, H. Chen, M. Su, Q. He, Z. Su: Financial Interests, Personal, Full or part-time Employment: Singlera Genomics Inc. R. Liu: Financial Interests, Personal, Officer: Singlera Genomics Inc. All other authors have declared no conflicts of interest.