Abstract / Summary
Abstract While anti-VEGF therapy is hypothesised to induce mesenchymal transition in high-grade glioma via hypoxia and myeloid remodelling, this dynamic remains unresolved at the resolution of defined malignant cell states in paediatric disease. Here, we reanalysed a longitudinal paediatric high-grade glioma single-nucleus atlas (Sussman et al., 2026; 401,253 cells, 16 patients, 32 samples) integrated with per-patient treatment histories. Applying multiple independent trajectory methods—including potency ordering, expression entropy, root-free Markov kernels, and optimal transport—we identified robust associations among malignant states. Although the original atlas grouped Transition 1 and Transition 2 as equivalent intermediate states between the astrocytic/mesenchymal and proneural programmes, our analyses reveal them as distinct entities with near-disjoint fate profiles and opposing positions across all lateness orderings. This confirms, via independent methods, a separation previously implied by motif-accessibility and ligand-receptor analyses. Crucially, Transition 2 represents a transcriptionally active state adjacent to the mesenchymal-like (MES-like) phenotype, accounting for 32.3% of MES-like nearest neighbours (a 2.9-fold enrichment over background). While initial reports noted no significant post-therapy shifts in overall malignant state composition, we found that the MES-like state contracts in bevacizumab-naive patients but expands under bevacizumab exposure, with the steepest increases observed following bevacizumab combined with checkpoint inhibition (0.07 to 0.372; patient-level permutation p = 0.027 for deltas, p = 0.003 for late-stage values). Concurrently, the myeloid compartment shifts toward a bone-marrow-derived macrophage (BMDM) phenotype. We demonstrate that BMDM polarisation strongly correlates with MES-like cellular proportions across samples (ρ = +0.83). This relationship remains robust after controlling for ambient RNA contamination, cell abundance, and leave-one-patient-out analyses, and persists after adjusting for hypoxia scores calculated outside the mesenchymal compartment (partial ρ = +0.64, p = 0.0008). Restricting our analysis to seven post-treatment samples from six bevacizumab-exposed patients revealed a striking cross-compartment coupling between myeloid VEGFA and malignant AXL (r = + 0.998 overall; r = + 0.946 specifically against AXL in MES-like cells). Within the malignant compartment, expression peaks in distinct states—AXL in Transition 2 (z = + 0.189) and VEGFA in the MES-like state (z = + 0.702). This intercellular interaction is arm-specific, invariant to leave-one-patient-out omission, resides at the 100th percentile of an empirical null distribution (n = 3,540 pairs), and withstands rigorous adjustment for hypoxia, cellular composition, sequencing depth, and primary technical or biological axes. This coupling is independent of canonical VEGF signalling: receptors VEGFR1, VEGFR2, and VEGFR3 are detectable in fewer than 1% to 8% of cells across both compartments, and tested candidate messengers fail to account for the axis. Neuropilins emerge as the sole abundant VEGFA-binding molecules (NRP1 expressed in 30% to 46% of cells). NRP1 is upregulated in bevacizumab-exposed tumours beyond compositional expectations and represents a cognate receptor spared by bevacizumab blockade. Finally, a malignant IL34-driven myeloid recruitment axis is evident in bevacizumab-treated tumours, though it remains perfectly collinear with concurrent checkpoint inhibition.