Abstract Liquid biopsy of cerebrospinal fluid (CSF) offers a minimally invasive alternative to surgical tissue sampling for diagnosing brain tumors or detection of minimal residual disease. Nanopore sequencing of cell-free DNA (cfDNA) is an increasingly used rapid method in molecular diagnostics. However, existing classification algorithms, originally designed for tissue samples, perform suboptimally with low tumor fractions typical for liquid biopsies. We developed NanoFlux, an ensemble model specifically designed to classify common brain tumors based on Nanopore-sequenced cfDNA isolated from the CSF. NanoFlux was trained on array-based methylation data from 3,871 tissue samples representing 18 tumor classes and controls, validated on a cohort of 140 Nanopore-sequenced cfDNA samples, and benchmarked against three established methods. Specific confidence thresholds for liquid biopsy applications were defined for all algorithms. In the validation cohort, NanoFlux correctly diagnosed 78/140 (55.7%) samples without any tumor misdiagnoses, outperforming other methods, which achieved at most 56/140 (40%) correct classifications. NanoFlux was further evaluated on an independent, real-world clinical cohort of 74 patients with imaging-suspected tumors who underwent tissue biopsies. Pre-operative detection of tumor DNA was possible in 20/60 (33.3%) patients with confirmed tumors. Correct tumor class prediction was feasible in 13/60 (21.7%) cases, compared to a maximum of 5 cases (8.3%) identified by alternative methods. In 14 cases, MRI-based tumor suspicion was rejected after tissue biopsy, and all such cases had correctly been foreseen to lack tumor DNA in pre-operative CSF. In summary, NanoFlux enables rapid and reliable CSF-based diagnostics, supporting the translation of CSF-based liquid biopsies into routine clinical practice.
Freese et al. (Tue,) studied this question.