916P Deep learning-based multimodal ensemble algorithm for multi-cancer detection and classification using cf-WGS

Abstract

Various cell-free DNA (cfDNA) features have been investigated for their potential use in early cancer detection and cancer type classification. In this study, we developed a multimodal ensemble algorithm which integrates two independent deep learning models in which one uses regional mutational density (RMD) and mutational signature as a training feature and the other uses cfDNA fragment end motif frequency and size (FEMs) profile. Cell-free whole genome sequencing data was generated from 1,088 patients (stage I: 31.9%, II: 23.8%, III: 17.5%, IV: 11.6%, unknown: 15.2%) with breast (n=418), ovarian (n=130), pancreatic (n=97), lung (n=92), esophageal (n=142), liver (n=157), and colon cancer (n=52) and 728 healthy controls. Sequence data was produced on average of 80 million reads using Novaseq 6000. Training (60%), validation (20%) and test (20%) dataset were split stratifying cancer type and stages. RMD and FEMS models were trained using multilayer perceptron and convolutional neural networks respectively. The average value of predicted probabilities from each model was used as the final ensemble prediction. Cancer detection performance reached an accuracy of 92.8% [95% confidence interval (CI): 90.1% to 95.2%] and an AUC of 0.979 (CI: 0.964 to 0.990) in the test dataset which consist of 229 cancer patients and 146 healthy controls. Cancer detection sensitivity was 91.7% (CI: 86.0% to 96.5%) at a specificity of 95%. In multi-cancer classification, the top 1 and 2 accuracy were 78.2% (CI: 77.9% to 78.2%) and 89.5% (CI: 89.3% to 89.6%) respectively. We developed a deep learning-based multimodal ensemble algorithm for multi-cancer detection and classification. Integrating a variety of cfDNA features all together enabled highly sensitive and accurate prediction. This result shows a great potential of cfDNA WGS-based liquid biopsy in general population multi-cancer screening.