Brain tumor typing during surgery drops its 2018 labels: 85 of 94 pediatric cases called correctly with confidence

Intraoperative sequencing yields only a sparse methylation profile, and the classifier must name the tumor type before sequencing finishes; Sturgeon V2 widens the reference data behind that call from a single 2018 dataset to nearly 10,000 samples. Publicly available sparse-profile classifiers all relied on the 2,801 samples and classification system published by Capper et al. in 2018, and the one classifier trained on a larger, newer dataset, MNP-FLEX, was reachable only through a web interface. In 94 retrospective pediatric intraoperative cases, V2 was confidently correct in 85, broadly correct in 2 and incorrect in 1, against 78, 3 and 2 for V1; across 1,880 simulated sequencing runs the error rate among confident calls fell from 7.1% to 2.8%.

Methylation profiles from different sources carry incompatible label systems, and V2 harmonizes them into one 102-class intraoperative taxonomy through iterative semi-supervised relabeling. Naive self-training at scale on the relabeled data lowered PR-AUC on the external nanopore benchmark from 0.932 to 0.900, showing that simple retraining was insufficient. Two concordance filters left 9,287 training samples; adding 620 relabeled samples retained from the CrossNN cohort brought the final training set to 9,907 and raised external benchmark PR-AUC from 0.939 to 0.948.

The architecture now reports the class with the highest mean score across 10 submodels and calibrates confidence in coverage bins, which suppresses class flipping as sequencing accumulates. V1 reported the most confident of four subclassifiers, and on difficult samples the predicted class could flip, leaving the final intraoperative call ambiguous. On 24 deeply sequenced nanopore samples downsampled to virtual timepoints from 663 to 25,000 CpG sites, V2 reached 87.9% recall across all timepoints with fewer than 0.1% errors.

The classifier targets sparse intraoperative methylation profiles, with classes defined at the granularity that changes intraoperative decisions, 102 in total, including meningioma subtypes, PATZ-fusion neuroepithelial tumors and PLAGL1/2-amplified embryonal tumors. Validation covers an independently labeled external nanopore benchmark, downsampled profiles from 177 untrained array samples, 24 deeply sequenced nanopore samples, 94 retrospective pediatric intraoperative cases and 1,880 simulated runs, plus 43 intraoperatively sequenced meningiomas used to compare four meningioma schemes. The software runs locally for non-commercial use, with model weights released.

Naive self-training at scale degraded performance through confirmation bias and hard-labeling of low-quality samples, while overly strict filtering yields underconfident classifiers that only classify the clearest cases. V2 does not resolve some finer molecular subtypes, including selected pediatric-type diffuse high-grade gliomas, medulloblastoma groups 3/4 and pineoblastoma, where comprehensive classifiers such as Heidelberg or Bethesda remain complementary. Caution is advised on low-purity samples, particularly array data, and direct prediction of the integrated meningioma risk score performed markedly worse, since the methylome alone does not reproduce that composite score. The results come from retrospective cohorts and simulated sequencing, pending prospective validation.