Deep learning–based cfDNA fragment–level analysis: A genome-scanning approach for detection of colorectal cancer.

243 Background: Cell-free DNA (cfDNA) is a promising biomarker for cancer detection. Circulating tumor DNA (ctDNA) in cancer patients exhibits unique features, such as mutation and fragmentomics, that differentiate it from cfDNA in healthy individuals. However, previous studies often analyze these features independently or without distinguishing cfDNA sources, which can lead to false positives or negatives. We aimed to develop a deep-learning method for cancer detection by combining genomic and epigenomic features at the fragment level. Methods: We analyzed cfDNA from 768 healthy individuals and 150 colon cancer patients (stages I–IV). The training set (n=605: 512 healthy, 93 patients) was used for 5-fold cross-validation, and the independent test set (n=313: 256 healthy, 57 patients) was used for evaluation. We selected markers from CpG-dense regions that are unmethylated in healthy cfDNA but tend to be hypermethylated in colon cancer tissues. For each marker, we constructed a multi-channel 2D tensor as input for our deep-learning model. We stitched raw sequencing reads and extracted six strand-specific channels: A, T, G, C nucleobase sequences, binary methylation status, and fragment alignment characteristics. A 2D convolutional neural network (CNN) was trained on these tensors. Individual marker scores were aggregated to produce a final cancer probability score. We benchmarked our model against conventional methods—Average Methylation Fraction (AMF), Depth profile, Fragment Size Ratio (FSR), and end-motif approach—using the mean per-marker Area Under the Curve (AUC). Results: Our method achieved a higher mean per-marker AUC (0.769 ± 0.039) than AMF (0.724 ± 0.051), Depth profile (0.528 ± 0.053), FSR (0.588 ± 0.042), and end-motif (0.566 ± 0.055). The final model achieved an AUC of 0.905 on the independent test set. It distinguished early-stage (I-II) and advanced-stage (III-IV) colorectal cancer with AUCs of 0.880 and 0.929, respectively. Conclusions: Integrating fragment-level cfDNA features provides high accuracy for colon cancer detection, especially for early-stage disease. Our genome-scanning approach allows for intuitive interpretation by visualizing cancer-associated genomic regions, improving the interpretability of deep-learning-based biomarker research and showing promise for more precise liquid biopsy tools.