Abstract 485P
Background
Early screening improves the survival of cancers. Here, we developed blood-based multi-cancers early detection (MCED) and cancer signal origin (CSO) approaches in 8 types of cancers including colorectal, esophageal, stomach, liver, pancreatic, lung, breast, and ovarian cancer.
Methods
A targeted methylation panel of 155,362 CpG sites (4000+ blocks) was constructed based on over 8,000 tissues and plasma methylation data (by internal whole-methylome-seq and public 450K data). We designed an case-control study (INSPECTOR) and recruited 1653 subjects. Training (total 1071; 538 cases) and validation sets (total 582; 312 cases) were randomly divided. cell-free DNA (cfDNA) was isolated from 10 ml peripheral blood and sequenced (median effective depth: over 1,000×). The machine learning MCED and CSO models were developed and validated in this study.
Results
803 non-cancer individuals as control and 850 cancer patients (stage I 22.4%, II 19.1%) were enrolled including esophageal (105), stomach (115), colorectal (113), liver (96), pancreatic (92), lung (133), breast (104), and ovarian cancer (92). Our MCED model showed 99.1% specificity and total sensitivity 70.3% (95% CI: 66.2% to 74.1%) in training set. In validations, specificity and total sensitivity were 99.3% and 72.8%, respectively, and the sensitivity were 49.3% for stage I, 66.0% for stage II, 78.4% for stage III, and 91.0% for stage IV. Notably, for stage I cancers, the sensitivity of esophageal (total: 71.4%), stomach(total: 77.5%), colorectal(total: 67.5%), liver(total: 100%), pancreatic(total: 64.5%), lung(total: 56.1%), breast(total: 45.7%), and ovarian cancer(total: 78.1%) were 57.1%, 60.0%, 36.4%, 100%, 66.7%, 23.5%, 33.3%, and 50.0%, respectively. Additionally, we selected a wide-ranging pathological subtype of cancers, including small cell lung cancer (sensitivity=28/29) and triple negative breast cancer (14/25). For CSO model, the accuracy of Top1 and Top2 was 90.5% and 95.5% in the training set, 83.7% and 91.2% in the validation set.
Conclusions
Our model exhibits remarkable results in identifying cancer signal and detecting the eight most common types of cancer.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.