Fragle: Deep Learning Model for Non-invasive ctDNA Cancer Detection - Report Summary

Monitoring cancer progression and treatment response in a non-invasive manner is a significant goal in oncology. Analyzing circulating tumor DNA (ctDNA) in the blood has emerged as a promising alternative to invasive biopsies. A new study introduces an innovative deep learning model called "Fragle" that enables accurate quantification of ctDNA from the fragment length density distribution of cell-free DNA (cfDNA).

The Fragle model is designed to learn the distinctive fragment length patterns of ctDNA compared to healthy cfDNA. It was trained on an extensive dataset of low-pass whole-genome sequencing (WGS) data from various cancer types and healthy control cohorts.

Validation demonstrated that Fragle outperformed simpler methods, achieving higher accuracy and lower detection limits. This improved sensitivity is crucial for detecting minimal residual disease (MRD) and early recurrence. Fragle is also compatible with targeted sequencing data, enhancing its potential clinical applicability.

Longitudinal analysis of plasma samples from colorectal cancer patients revealed a strong concordance between ctDNA dynamics quantified by Fragle and treatment responses. In patients with resected lung cancer, Fragle outperformed simpler gene panels in predicting MRD and risk stratification.

Fragle's versatility, speed, and accuracy make it a promising tool for various clinical applications, including early cancer detection, treatment monitoring, and MRD diagnosis. Its non-invasive nature also holds promise for cancer screening and drug development.

In conclusion, Fragle represents a significant step forward in non-invasive cancer monitoring, with the potential to improve cancer management and patient outcomes.

Source: A deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengths