AI-Assisted Molecular Review Refines SMO Inhibitor Candidacy in SHH Medulloblastoma

AbstractMedulloblastoma is a rare adult central nervous system tumor, most often activated by the sonic hedgehog (SHH) signaling pathway. Standard treatment is maximal safe resection, craniospinal irradiation, and multiagent chemotherapy. No agent has an established maintenance role in any risk group. Smoothened (SMO) inhibitors, which target the SHH pathway, are a candidate, but their use is off-label (i.e., not approved by the U.S. Food and Drug Administration for medulloblastoma), and their optimal timing - or whether they are warranted at all - is unresolved. No randomized trial can settle this, given the rarity of specific genotype-risk combinations in adult medulloblastoma.We present the case of a 40-year-old woman with unresectable, SHH-activated medulloblastoma who received standard chemoradiation. Posttreatment imaging was equivocal. NanoString classified the tumor as WNT, conflicting with methylation and pathology, both of which supported SHH activation. As was later established, a co-occurring TP53 mutation placed her in the very-high-risk category.We used Claude (Anthropic), configured with a custom neuro-oncology instruction set, to integrate pathology, methylation, NanoString, and next-generation sequencing with clinical and imaging data. This was verified against readings by our team's pathologist and neuroradiologist. The synthesis resolved the discordance by weighting methylation as the gold standard, corroborated by GAB1 and beta-catenin staining at the protein level. It identified a mechanistic explanation for the NanoString result and flagged a low-confidence methylation subtype. Next-generation sequencing showed no PTCH1, SMO, or SUFU mutation; weighed against the tumor's broader molecular context, this placed it in a genotype-indeterminate category for SMO inhibitor response - absence of evidence for benefit, not evidence of absent benefit. No SMO inhibitor was given; we chose close surveillance instead. The patient remains in remission, with an 11-month progression-free survival at last follow-up.To assess whether this synthesis depended on a single model, we repeated the analysis with two additional large language models, Gemini and GPT-5, using identical source reports and instructions. Both reached the same subgroup classification and SMO inhibitor candidacy conclusion, though agreement on the underlying discordance mechanism reasoning was only partial, indicating the conclusion was not tied to any one model's reasoning.