PATH-33. Comparison of Nanopore Classification Pipelines and Basecalling Models for Solid and Liquid Tumor Biopsies

Abstract Nanopore sequencing of CNS tumors enables real-time DNA methylation-based tumor type prediction and is revolutionizing diagnostic workflows. Multiple classifier tools have been developed, which differ in machine learning frameworks, which we sought to benchmark here. We benchmarked three pipelines, MethyLYZR (naïve Bayes model), Sturgeon (transfer-learned neural network), and a custom-adapted NanoDx (crossNN neural network), and three basecalling models, which differ in computational demands (fast, high-accuracy [hac], and super-accuracy [sup]). All classifiers were run on hac basecalled data, while for nanoDx the three basecalling models were compared. DNA was isolated from 12 samples (6 CSF and 6 fresh tumor biopsies) using the rapid (n=4) or ligation sequencing protocols (n=8), and sequenced for 155 to 3,789 minutes (mean coverage: 0.69×). Eight of 12 samples yielded classifiable results consistent with the histology (4/6 liquids, 4/6 solids). Among classifiable cases, concordance rate was 100%, whereas thresholds (MethYLYZR 0.6, Sturgeon 0.8, nanoDx 0.2) were not met in 3 cases, and models hallucinated in case of the CSF of a pancreatic neuroendocrine cancer patient. The minimum required DNA input was 0.63 fmol/µL and did not correlate with CpG yield (Spearman’s ρ = –0.69, p = 0.15). Comparison of the basecalling models yielded a mean of 165,213 CpGs for fast, 160,310 for hac, and 162,858 for sup. Predictive scores increased on average by 11.47% from fast to hac and 9.44% from hac to sup, while computational time on a GPU-HPC was 55.4 minutes for fast, 74.75 minutes for hac, and 214 minutes for sup. Excellent concordance among the three classification pipelines underscores their analytical consistency, while the high-accuracy (hac) basecalling model emerges as a pragmatic compromise between accuracy, number of CpGs detected, and computational time investment for reliable CNS tumor classification.