Convergent Two-Classifier Transcriptomic Stratification of Five Antimeningioma Drug-Target Programs Across Public Multi-Omic Meningioma Cohorts

1. Introduction

Meningiomas are the most common primary intracranial tumour, yet only three systemic agents have shown meaningful activity in prospective trials [1,2]. The Heidelberg meningioma group has recently substantially expanded the pharmacological toolbox: Jungwirth et al. screened 119 FDA-approved drugs on meningioma cell lines (Clin Cancer Res 2023; hereafter CCR 2023) [3] and identified carfilzomib, omacetaxine, ixabepilone, and romidepsin as the four most potent compounds (IC₅₀ 0.12–9.5 nmol/L, largely via G2–M arrest and apoptosis). Jungwirth et al. then extended this to 60 molecularly characterised patient-derived tumour organoids (Sci Transl Med 2026; hereafter StM 2026) [4], identifying panobinostat as the lead pan-HDAC inhibitor across in vitro, ex vivo, and in vivo models, and tracing resistance to an HDAC8—*TGFβ—*EMT axis whose genetic ablation restores sensitivity.

Both screens are rigorous pharmacological experiments, yet both share the standard limitation of functional pharmacogenomic studies: compound activity is reported at the cohort level, not stratified by any now-validated meningioma molecular subgroup system [5,6,7]. There is no patient-selection layer handable back to the clinic. Two mature, independently derived classification systems are publicly available for meningioma. Nassiri et al. (Nature 2021) [5] profiled 185 meningiomas by methylation, WES, bulk mRNA, and snRNA-seq, defining four mutually consistent groups (MG1 Immunogenic / MG2 benign NF2-wildtype / MG3 hypermetabolic / MG4 proliferative) with processed data deposited under cBioPortal mng_utoronto_2021. Choudhury and the Bi/Raleigh lab (Nature Genetics 2022, refined Neuro Oncology 2023) [6,7] integrated methylation, genetics, transcriptomics, proteomics, and single-cell data on 565 meningiomas to define a complementary three-tier system (Merlin-intact / Immune-enriched / Hypermitotic) framed around therapeutic vulnerabilities, with paired RNA-seq for the N = 185 discovery subset available under GEO GSE183653.

The existence of two methodologically distinct, publicly accessible classification systems creates a rare methodological opportunity: if a target-gene program is differentially expressed across subgroups in both classifiers, the signal is proportionally less likely to reflect classifier-specific artefact. This two-classifier

convergence logic—combined with explicit falsifiability testing of the cross-study subgroup mapping via exhaustive pairing enumeration and honest declaration of currently non-evaluable hypotheses—is the core methodological contribution of the present work.

1.1. Pre-Specified Hypotheses (Locked Before Any Statistical Testing)

H1 (classifier concordance).

The literature-registered mapping between Nassiri 4-group labels and Bi-lab 3-group labels (MG1 - Immune-enriched, MG2 - Merlin-intact, MG3∪ MG4 - Hypermitotic) achieves statistically significant cross-study agreement as quantified by Cohen'sκ on the projected 3×3 shared label space, and the true mapping ranks strictly above all 11 alternative many-to-one 4—*3 pairings on bothκ and raw agreement (pairing-specificity test).

H2 (target-program stratification).

Expression of the five mechanism-matched drug-target programs and the StM-2026-specific HDAC8—*TGFβ—*EMT resistance program differs significantly across Nassiri molecular subgroups after Benjamini–Hochberg FDR (BH-FDR; α = 0.10) applied across the full program × subgroup family, and this difference is not explained by WHO grade alone (grade-adjusted conditional F-test).

H3 (NF2 CNA-loss proxy association).

Target-program z-scores differ between samples classified as NF2 Intact vs NF2 Mutant/Loss by a GISTIC-derived copy-number-loss proxy (nf2_cna_loss_proxy), with BH FDR at α = 0.10 across the five-program H3 family. The proxy is explicitly labelled a structural proxy, not a definitive biallelic-inactivation call (Methods and Limitations).

H4 (external replication).

Directional concordance of subgroup × program associations discovered in the Nassiri discovery cohort will be assessed in at least one independent public expression cohort with verified subgroup metadata. Pre-specified contingency: if no independent cohort carries both expression and verified subgroup labels, H4 is declared not evaluable rather than imputed.

H5 (exploratory recurrence-free survival, hypothesis-generating only).

Where and only where ≥ 20 first-recurrence events with a time-to-event variable are available, a grade- and NF2-adjusted Cox PH model of RFS on program-score high/low median split is pre-specified. H5 is never used to support primary causal claims; if event/time metadata are not deposited, H5 is declared not evaluable.

2. Results

2.1. Aim 0: Cohort Data Completeness and Hypothesis Evaluability

Cohort accessions and metadata completeness were locked before any statistical testing (Aim 0; results/tables/aim0_cohort_lock.*). The Nassiri 2021 discovery backbone comprised N = 121 samples with complete 4-group labels, WHO grade, and nf2_cna_loss_proxy variables on cBioPortal mng_utoronto_2021 (121/185 paper samples available on the processed open-access portal entry). Group distribution: Immunogenic 17, MG2 32, hypermetabolic 43, proliferative 29. nf2_cna_loss_proxy: Intact 37 (30.6%), Mutant/Loss 84 (69.4%). Critically, the Immunogenic (MG1) subset showed 0/17 NF2-Intact calls (100% CNA-loss proxy), directionally consistent with the literature claim of “invariable biallelic NF2 inactivation” for this group, while MG2 showed 27/32 (84.4%) Intact, consistent with the NF2-wildtype benign designation [5,8].

Three replication GEO cohorts (GSE136661 N = 160; GSE77259; GSE94474) carried WHO grade and age/sex metadata but 0% subgroup and 0% NF2-status coverage on GSM-level SOFT inspection; Bi-lab GSE212666 (N = 302 validation RNA-seq) carried 0% phenotype metadata (tissue only). H4 and H5 were therefore declared not evaluable pre-analysis per Aim 0 lock (full status: Supp Table S0 / aim0_cohort_status_report.txt).

2.2. H1 — Classifier Concordance and Pairing-Specificity Falsifiability

The Bi-lab 3-group labels for Nassiri samples were derived via a Ridge-penalised multinomial bridging classifier trained on the GSE183653 N = 185 Bi-lab discovery TPM matrix (2,000 most variably expressed genes, standardised on training set only; Methods). Leave-one-out cross-validation (LOOCV) on the training set yielded accuracy 93.0% (κ = 0.894), with per-class recall: Merlin-intact 0.958, Immune-enriched 0.867, Hypermitotic 0.962 (Supp cv_metrics_ridge.csv). Predicted 3-group distribution on Nassiri N = 121: Hypermitotic 45, Immune-enriched 37, Merlin-intact 39.

Raw 4×3 cross-tabulation (Nassiri native groups × predicted Bi groups) is shown in Table 1 (full counts and row-percentages in Supp Table S1). Raw association on the full 4×3 space: χ²(6 df) = 99.58, p = < 1 × 10^-16, Cramér's V = 0.641, confirming strong cross-structure dependence before any label projection.

Mapped 3×3 Cohen’s κ: after projecting Nassiri labels onto the Bi label space via the pre-registered mapping {Immunogenic → Immune-enriched, MG2 → Merlin-intact, hypermetabolic → Hypermitotic, proliferative → Hypermitotic}, agreement on the resulting 3×3 shared-space table was moderate-to substantial and highly significant: κ = 0.568 (95% CI computed via statsmodels cohens_kappa; asymptotic p = < 1 × 10^-16; Supp Table S1).

Pairing-specificity test. To rule out the trivial explanation that “any 4→3 pairing works”, the true mapping was scored against the full enumeration of 11 alternative pairings that also route exactly two Nassiri groups into the Hypermitotic bin and the remaining two as singletons (6 choices of merge-pair × 2! singleton permutations = 12 total; Table 2, Supp Table S1b). The true mapping ranked #1 of 12 by both κ and raw agreement (71.9%). Key contrasts against adversarial alternatives:

• Δκ vs the best-performing wrong pairing = +0.302 (true κ = 0.568 vs best-alt κ = 0.266).
• Singleton-swap placebo (merge still correct: MG3∪MG4→Hypermitotic, but singletons swapped: MG1↔Merlin-intact, MG2↔Immune-enriched) collapsed to κ = 0.050, confirming the singleton identity is non-trivially recovered.
• Worst wrong-merge choice (MG1∪MG2 pooled into Hypermitotic) gave κ = -0.324, confirming that the literature's choice of which two groups pool is empirically discriminable.
• Pre-specified pass criteria were met: pass_strict = True (rank #1 on both κ + agreement) and pass_any_alt = True (κ > all 11 alternatives).

H1 is satisfied at both the nominal-significance and the pairing-specificity falsifiability levels.

2.3. H2 — Subgroup Stratification (Nassiri 4-Group, Native Labels)

Program scores were computed via rank-based ssGSEA (α = 0.25) [9,10] and z-transformed per program. Kruskal–Wallis omnibus tests across the four Nassiri groups, with BH-FDR applied across the full 5- program H2 family (α = 0.10), rejected for all five programs (Table 3):

Pairwise Mann–Whitney U tests with per-program BH-FDR (full results in Supp Table S2a) highlighted three reproducible structure features, consistent with the subgroup biology:

• MG3/MG4 (Hypermitotic pool) asymmetry. The hypermetabolic (MG3) vs proliferative (MG4) contrast was the weakest pairwise comparison for all five programs (all pairwise pBH > 0.10 after per-program BH-FDR; range: HDAC pBH = 0.313 to Tubulin pBH = 0.731), independently justifying the literature pooling of MG3∪MG4 into a single Hypermitotic bin.
• Immunogenic (MG1) vs proliferative (MG4) — HDAC and HDAC8-resistance extremes. The Immunogenic→proliferative step was the single largest effect for the HDAC target program (Δ median z = -1.475, Cliff's δ = -0.761, family-wide pBH = 6.3 × 10^-4) and for the StM 2026 HDAC8/TGFβ/EMT resistance program (Δ = -1.332, δ = -0.582, pBH = 0.004).
• MG2 (Merlin-intact) separation — strongest translation and HDAC8-resistance contrasts. Immunogenic↔MG2 was the strongest pairwise effect for the translation-elongation (omacetaxine) program: Δ median z = -1.270 (MG2 minus Immunogenic), Cliff's δ = -0.673, family-wide pBH = 0.001. The same MG2↔Immunogenic contrast was also the strongest