mFrag: Integrating External DNA Methylation References Improves Lung Cancer Early Detection by Fragmentomics

Fragmentomics is a method that examines features of circulating cell-free DNA fragments, such as fragment length or end motifs, obtained from whole-genome sequencing of liquid biopsy samples like blood or other bodily fluids. This approach enables non-invasive early cancer detection. In this study, researchers introduce mFrag, a framework that combines external methylation reference atlases with fragmentomic features to enhance lung cancer detection. mFrag consists of two complementary methods: mFrag-Weight, which assigns weights to fragmentomic features using DNA methylation information, and mFrag-Net, which merges fragmentomic features and external DNA methylation data via neural networks. The final prediction score is derived by averaging the output probabilities of the two methods.

To evaluate the efficacy of mFrag, the researchers trained the model using data from 287 individuals at risk for lung cancer. They then tested the model in an independent case-control setting comprising 385 non-cancer controls and 46 cancer patients (28 at stage I, 12 at stage II, and 6 at stage III or IV). The results demonstrated that mFrag significantly outperformed a DELFI-style fragmentomic approach, raising the probability of accurate classification (PR-AUC) from 0.455 to 0.720 and sensitivity at 99% specificity from 0.043 to 0.478, with ROC-AUC increasing from 0.871 to 0.908.

The improvements in mFrag's performance were particularly notable in detecting stage I lung cancer. PR-AUC rose from 0.314 to 0.558, and ROC-AUC increased from 0.855 to 0.894. When applied to end 4-mer motif features, mFrag continued to outperform an unweighted baseline, with the gains attributed to both mFrag-Weight and mFrag-Net.

Overall, this study highlights the potential of mFrag in improving the accuracy of fragmentomics-based lung cancer detection by integrating external DNA methylation data and leveraging the complementary strengths of its two component models, all without incurring any additional data-generating costs.