Tumor-naïve ctDNA detection with deep learning-enhanced error suppression for sensitive mutation calling.

BACKGROUND

Circulating tumor DNA (ctDNA) detection offers minimally invasive monitoring of cancer from blood samples. While tumor-informed approaches are sensitive, their clinical application is limited by cost, tissue availability, and turnaround time. In contrast, tumor-naïve assays offer faster turnaround times and broader applicability, but often sacrifice sensitivity.

METHODS

Here, we developed DEEPctMUT, a computational pipeline for tumor-naïve ctDNA detection that integrates unique molecular identifiers (UMIs) with three complementary error-polishing strategies: (1) machine learning-based sequencing error suppression, (2) deep learning-based background noise filtering (DeepES), and (3) removal of clonal hematopoietic and germline variants using patient matched PBMC. We designed an efficient sequencing panel and developed and validated DEEPctMUT using cell lines, spike-in samples, healthy, and CRC plasma samples.

RESULTS

DEEPctMUT accurately detected mutations down to 0.03% variant allele frequency (VAF), outperforming other tumor-naïve methods. In a head-to-head comparison, our pipeline identified pre-surgical CRC cases with 100% sensitivity, whereas the Roche Avenio Surveillance Kit achieved only 50% sensitivity. Furthermore, a panel-independent version of DEEPctMUT could improve the performance of Roche Avenio data without additional assay-specific training. We provide an extensive ground-truth dataset and make DEEPctMUT, including all trained models, available as an easy-to-use Nextflow pipeline.

CONCLUSIONS

By using advanced computational error-polishing techniques, DEEPctMUT can substantially reduce technical artifacts, allowing our tumor-naïve method to achieve a sensitivity comparable to tumor-informed approaches.