Single-cell geometry of drug responses across malignant glioma states

Malignant gliomas exhibit substantial transcriptional heterogeneity: AC-, OPC-, NPC- and MES-like states can coexist within the same tumor, and a drug affecting one state may leave others relatively unaffected. Whether these states respond in a shared direction, or diverge, is rarely tested directly, since drug response is usually summarized by one score that collapses this heterogeneity. A public single-cell RNA-sequencing resource of drug-treated glioma slice cultures (391,444 cells) is reanalyzed here to test this directly. After restricting to 252,106 transformed glioma cells, 43 treatment-vehicle comparisons from 19 TissueID-defined units are evaluated for response-direction coherence (C, similarity of state-conditioned treatment vectors) and change in state separation (K, state-centroid contraction or expansion). C proves to be the reproducible signal: it survives high-confidence state filtering (r=0.931), an independently selected 500-gene space (r=0.634), the published Neftel modules (r=0.783) and non-overlapping gene cross-fitting (r=0.809), and exceeds two null models in 41 of 43 comparisons, a result that also holds across the 19 biological units themselves (19 of 19 positive; P=3.8e-6). Within-drug coherence exceeds between-drug coherence in all 52 drug pairs and all 11 multi-drug units (P=0.000488): response direction is drug-specific, not generic. K tells a different story: stable only when the feature space is fixed, it loses concordance whenever that space shifts, with only about half of comparisons separating from the null. Malignant glioma states largely agree on the direction of drug response even when they disagree on its magnitude.