Neftel et al. profiled 28 glioblastomas (20 adult, 8 pediatric) with SMART-Seq2, yielding 24,131 profiled cells of which 7,930 passed quality control (~5,730 genes/cell), plus a 10x Genomics validation set of 16,201 cells from 9 tumors and 401 TCGA bulk samples. They converged intra-tumoral variability on six meta-modules (39β50 genes each) defining NPC-like, OPC-like, AC-like, and MES-like malignant states, with MES1/MES2 and NPC1/NPC2 subprograms .
The decisive mechanistic findings: each tumor contains β₯2 states; 15% of cells are hybrids; 37 CNA-defined subclones across 12 tumors each contained multiple states (same-state pairs 51% within vs 46% across subclones), arguing genetic subclones do not dictate states. EGFR amplification associated with AC-like, PDGFRA with OPC-like, CDK4 with NPC-like (all p<10β5 in TCGA), and 5q loss/NF1 with MES-like states . Functional validation: EGFR overexpression drove AC-like programs and GFAP+ induction in mouse NPCs; CDK4 drove NPC-like programs and NPC proliferation (p<0.0001). Barcoded lineage tracing showed 39% of multi-cell barcodes spanning multiple states, and single barcodes giving rise to all four statesβdirect in vivo evidence of plasticity .
Strengths: Multiple orthogonal validations (RNA-ISH in 10 specimens, cross-platform 10x replication, mouse models, human PDXs, TCGA bulk integration) and open data (SCP393, GSE131928). Limitations the authors partially acknowledge: CNA inference from scRNA-seq misses small events; barcoded mouse models may not recapitulate human biology; microenvironment influences incompletely captured .
Independent follow-up qualifies the model: GLICO cerebral organoid experiments confirmed that microenvironment is necessary to maintain primary-tumor state distributionsβpurely cell-intrinsic plasticity is not the whole story . A 2025 preprint extended the framework clinically: drug-induced reversible state shifts (panobinostatβMES-like) drive resistance, and state-selective lethality combinations (panobinostat+paxalisib, CI<1) exploit the plasticity Neftel described .
Remaining gaps: No direct causal test that manipulating EGFR/CDK4/PDGFRA in vivo shifts the steady-state distribution; MES-like state had no significant CNA association in TCGA amplification analysis (only 5q loss/NF1 associations); temporal evolution of states during therapy (Stupp regimen) untested; cell-state assignments from ~400-gene signatures may conflate state with stress responses (MES2 is explicitly hypoxia-linked, a possible dissociation artifact). What would falsify the model: subclones strictly mapping to single states, or failure of sorted single-state cells to regenerate multi-state tumors.
Note: the second figure encodes qualitative association categories reported in the paper (all p-values stated as <10β5 or significant by hypergeometric test), not measured effect sizes.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.