Society: AGA

Background
Colorectal cancer (CRC) remains a leading cause of cancer related mortality worldwide. We utilized cell-free DNA (cfDNA) methylation, fragmentation characteristics of selected cancer-related biomarker regions, tumor-derived signal deduction and a machine learning algorithm to refine a blood test for the early detection of CRC and advanced adenomas (AA). The aim of the study was to assess the diagnostic accuracy of the test for CRC.
Methods
This was a prospective, international (Spain, Ukraine, Germany and USA [part of NCT04792684 study] population), observational cohort study. Plasma samples from 997 patients were collected either prior to a scheduled screening colonoscopy or prior to colonic surgery for primary CRC. cfDNA samples from 170 early stage (I-II), 128 late-stage (III-IV) CRC patients (mean age 66 [44-84], female 48%, distal cancers 60%), 149 AA patients (63 high grade dysplasia; 84 low grade, > 1cm) and 550 age, gender and country of origin matched colonoscopy-checked controls were included. 155 of the control patients had a negative colonoscopy finding (cNEG), 337 had benign findings of diverticulosis, hemorrhoids, previously undiagnosed gastrointestinal diseases and/or hyperplastic polyps (BEN), 58 had non-advanced adenomas (NAA). Samples were analyzed utilizing hybrid-capture based sequencing methodology. Panel of targeted biomarkers was previously identified through tissue- and plasma-based discovery and verification workflow. Individual cfDNA fragments belonging to each biomarker region were scored for cancer-specific methylation and fragmentation signals. Finally, calculated scores were used in prediction model building and testing for establishing panel accuracy.
Results
Prediction model utilizing a panel of methylation and fragmentation scores originating from biomarkers belonging to relevant cancer development and progression related pathways, correctly classified 93% (276/298) of CRC patients and 54% (81/149) AA patients. Sensitivity per cancer stage ranged from 85% (48/56) for stage I, 94% (107/114) stage II, 94% (90/96) stage III and 97% (31/32) stage IV. Fragmentation signals contributed most to early-stage cancers (I-II), while methylation signals were more significant for late stage (III-IV) detection. High grade dysplasia AA sensitivity was 52% (33/63), while low grade >1cm AA sensitivity was 57% (48/84). Specificity of the model was 92% (504/550), with 83% (48/58) NAA, 93% (312/337) BEN and 93% (144/155) cNEG patients correctly identified. Lesion location, gender, age, BMI and country of origin were not significantly (p> 0.05) correlated to prediction outcome.

Conclusions
Use of methylation and fragmentation characteristics of cancer-related cfDNA regions, combined with a machine-learning algorithm is highly accurate for early-stage (I-II) CRCs (91% sensitivity) and AA (54% sensitivity) at 92% specificity).

Marko Chersicola

Appearances