Fragmentomics powers improved classification of somatic mutations in liquid biopsy

Introduction: Accurate discrimination between somatic and germline variants in cell-free DNA (cfDNA) is critical for precision oncology. Somatic variant calling in liquid biopsy can be challenging, particularly for variants with high mutant allele frequency (MAF) in high tumor fraction samples. We developed a fragmentomics-based machine learning model to improve somatic variant classification in liquid biopsy. Methods: We benchmarked our model on 4,250 clinical samples from 3,313 unique patients processed on Guardant360 Liquid (Guardant Health, Palo Alto, CA), where variant germline status could be confidently derived from longitudinal data. Training data comprised 5,253 unique variants; independent test data included 11,612 variants. We engineered 55 fragmentomics features and multiple classifiers were evaluated; optimal performance was demonstrated with logistic regression with L2 penalty using fragment length and relative VAF features. We chose a cutoff threshold for classification scores to maintain less than 10% false positives by the model. The model was integrated as an additional correction step to re-evaluate variants classified as germline by a baseline calling algorithm. Results: Starting with a truth set of 26,900 variants from 2,874 unique samples whose somatic or germline origin were known from longitudinal data, 22,687 had a truly somatic origin and 4,213 were germline. Our baseline calling algorithm assigned germline status to 770/22,678 somatic variants and likelihood of mis-assignment was associated with a high ctDNA tumor fraction. The fragmentome-assisted approach was able to 'rescue' 558 of the mis-assigned variants including 226 with direct clinical actionability. In terms of sensitivity (true somatic variant/true somatic variant+somatic variant called germline) and specificity (true germline variant/true germline variant+germline variant called somatic), our approach improved sensitivity compared to the baseline calling algorithm from 96.61% to 99.07%, while specificity decreased from 99.00% to 93.38%. Despite the drop in specificity, accuracy and F1 score increased by 1.18% and 0.76%, respectively. The logistic regression model demonstrated robustness across cancer types and varying fragment profiles, with the most predictive features being di-nucleosome fragment lengths and relative allele frequencies over fragments ranging between 180 to 220 bases. Conclusions: Inclusion of model-based fragmentomics signals in a cfDNA germline-somatic discrimination caller significantly improves somatic variant detection sensitivity while maintaining high accuracy, enabling confident reporting of additional clinically actionable alterations. The boost in sensitivity at the expense of specificity is clinically acceptable given the intended use of the test for detection of targetable somatic alterations.