410P - Systematic evaluation of cell-free DNA fragmentation patterns for cancer diagnosis and enhanced cancer detection through integration of multiple fragmentations

Background

Cell-free DNA (cfDNA) fragmentation patterns hold immense potential for early cancer detection. However, the lack of systematic comparison among these patterns has impeded their broader research and practical implementation.

Methods

Here, we collected over 1,382 plasma cfDNA sequencing samples from diverse sources, covering eight cancer types including breast cancer, cholangiocarcinoma, colorectal cancer, gastric cancer, lung cancer, ovarian cancer, pancreatic cancer, and liver cancer. Considering that cfDNA within open chromatin regions is more susceptible to fragmentation, we leveraged ten fragmentation patterns within open chromatin regions as features and employed machine learning techniques to evaluate their performance. The considered fragmentation patterns included Windowed Protection Score, Preferred end coordinates, Coverage, Orientation-aware Cell-free Fragmentation, DNA Evaluation of Fragments for early Interception, Fragment Size Ratio, Fragment Size Distribution, End Motif preferences, Promoter Fragmentation Entropy, and Integrated Fragmentation Score.

Results

All fragmentation patterns demonstrated discernible classification capabilities, and the category of fragmentation patterns incorporating both fragment length and coverage information exhibited robust predictive capacities. The ensemble model integrating all these fragmentation patterns further improved performance in cancer detection and tissue-of-origin analysis. Biologically, crucial features of the model captured critical regulatory regions involved in cancer pathogenesis.

Conclusions

A comprehensive machine-learning-based evaluation of ten major cfDNA fragmentation patterns for early cancer detection was performed. Enhanced performance in cancer diagnosis and tissue-of-origin estimation was achieved, through integration of these fragmentation patterns in an ensemble model with biological interpretability.

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.