BGPT: Paper Review: A spatiotemporal cancer cell trajectory underlies glioblastoma heterogeneity

Explore by Goal

Quick Explanation Copied

Core claim (critical read)

Dataset scale: 12 human IDH-wt primary glioblastomas with multi-region single-nucleus multiome (snRNA-seq+snATAC-seq) plus spatial transcriptomics (Visium; some Xenium/IHC) and spatial WGS via LCM, totaling ~1.0M nuclei and ~338k Visium spot transcriptomes across 97 Visium sections.
Main trajectory: malignant states transition (putatively in “latent time”) from developmental-like (OPC/NPC-like, AC-prog-like) toward gliosis (AC-gliosis-like → gliosis) and then hypoxia, with spatial zonation across tumor anatomy.
Coupling to myeloid TME: the trajectory is claimed to be “regionalised” by distinct myeloid compartments and predicted myeloid→malignant ligand-receptor signaling programs (TAM classes with different niches across dev-like vs gliosis/hypoxia regions).

Most important skeptical question: this is largely inference from cross-sectional snapshots (RNA-velocity/fate modeling) and deconvolution-based spatial inference; the trajectory is not directly validated by lineage tracing or perturbations in vivo/human tissue.

Long Explanation

Paper review (evidence-based, skeptical)

A spatiotemporal cancer cell trajectory underlies glioblastoma heterogeneity

Preprint DOI: 10.1101/2025.05.13.653495 (May 14, 2025)

Key biological thesis

GBM heterogeneity is organized by a conserved spatiotemporal malignant trajectory (dev-like → gliosis → hypoxia) that is spatially zonated, shared across subclones, and coupled to regionalised myeloid niches.

Visuals (what the study did)

How the authors infer “trajectory” (mechanistic skepticism)

Trajectory method:

Infer malignant cell states by integrating snRNA-seq clustering (with CN gain chr7 / CN loss chr10 for malignancy enrichment) and annotate “dev-like”, “gliosis/hypoxia”, and “proliferative” classes.
Call subclones using multiome CNA inference and validate with DNA-level spatial WGS via LCM (and alleleIntegrator guided by LCM-WGS allele information).
Infer temporal ordering within each subclone via cell2fate (a Bayesian RNA-velocity/fate framework) on snRNA-seq spliced/unspliced counts, then detect recurrent dev-like → AC-gliosis/hypoxia trajectories.

Skeptical note: RNA velocity/cell-fate methods infer relative ordering from snapshot transcriptional dynamics; the paper itself acknowledges lack of lineage tracing/perturbation validation and does not test reversibility of gliosis/hypoxia back to dev-like states.

Main quantitative claims that can be stress-tested

Interpretation risk:

Selection bias within the inference: only subclones meeting robust/quality criteria (and sufficient cell/state diversity) are included in trajectory conclusions. The paper states 26/58 subclones met robust criteria; others were “poorly discernible” due to small numbers or low diversity.
Conflation of trajectory with zoning: the spatial zonation supports an axis of change, but spatial proximity/deconvolution can also reflect differential survival, cell mixing, or sampling/registration artifacts—not only state transition. The authors do use orthogonal Xenium and IHC to validate zonation of selected markers (e.g., AKAP12 and HILPDA).

Spatial organization into niches (and what is inferred vs measured)

Niche inference uses cell2location deconvolution on Visium spots, then NMF factorization to define tissue niches, and constructs a network from pairwise spatial proximities.

Critical blind spot: deconvolution methods are sensitive to reference signature completeness and to how well snRNA-seq states capture within-spot diversity; the paper attempts to mitigate this by using tumor-specific malignant reference signatures in cell2location training and by validating selected zonation using Xenium/IHC.

Clonal “shared origins” across dev-like and gliosis/hypoxia states

Claim: most subclones contain both dev-like and gliosis-hypoxia malignant cells, implying shared clonal origins along the trajectory.
Nuance / counterpoint: the paper reports “surprising” exceptions—some late-genomic alterations (chr17 alterations in specific tumours) were associated with strongly enriched cellular compositions, suggesting mutations can bias trajectory outcomes.

Quantifying clonal mixing in situ (SpaceTree)

SpaceTree is introduced as a joint clone + cell-state deconvolution model for Visium spots using a reference scRNA-seq dataset, implemented as a multi-task graph neural network with label propagation.

What would disprove it: if clonal entropy/mixture estimates are driven mainly by tumor-purity and reference signature correlations, rather than real subclone intermixing. The authors claim orthogonal validation by comparing SpaceTree domains to LCM SNV-based deconvolution on adjacent tissue sections.

Myeloid coupling to the malignant trajectory: supportive evidence + what remains inferential

Myeloid taxonomy: 11 myeloid subtypes across 3 phenotypic classes (resident/pro-inflammatory; infiltrating/anti-inflammatory; stress-response/angiogenic TAMs).
Spatial coupling: myeloid classes are enriched in different GB tissue niches (dev-like immune hotspots vs gliosis/hypoxia-associated TAM enrichment near PNZ/PAN/necrosis).
CCC inference: LIANA+ consensus across multiple ligand-receptor inference methods plus Tensor-cell2cell factorization identifies conserved myeloid ligands targeting dev-like vs gliosis/hypoxia malignant receiver states.

Major inferential limitation: CCC analyses infer signaling potential from expression and curated interaction databases; they do not prove ligand availability, receptor engagement, or functional downstream effects in vivo. The paper’s own limitations do not claim causal experimental validation.

Evidence-map (a compact visual logic graph)

How to read this: the paper’s strongest support is the multi-modal consistency of zoning and gene programs plus clone-aware intermixing. The weakest link is causality (trajectory inferred from snapshots; CCC inferred from expression/priors).

Bottom-line scientific critique (most actionable)

What looks strong

Multi-modal, multi-region design at scale (snRNA+snATAC multiome + Visium across many sections; some Xenium/IHC; plus spatial WGS).
Orthogonal support for spatial zoning (marker zonation validated using Xenium/IHC).
Clone-aware interpretation: trajectory interpretation is conditioned on subclones and validated with DNA-level LCM-WGS-informed allele/copy-number calling.

What is most uncertain

Trajectory ≠ demonstrated temporal causality: cell2fate/cell-fate inference provides posterior temporal ordering but does not show actual lineage transitions in vivo/human tissue over time.
Spatial deconvolution assumptions: mapping Visium spots to malignant states and quantifying clonal entropy depend on reference signatures; mis-specification can yield spurious “niche” structure or apparent intermixing. The paper attempts mitigation via tumor-specific reference signatures and orthogonal comparisons, but the residual risk remains.
CCC inference is potential-only: ligand-receptor interactions are computed from expression + curated databases + consensus methods; functional validation of predicted ligand-receptor effects on trajectory/state transitions is not provided.

Further BGPT actions (bespoke)

Author reviews on BGPT

Feedback:

Updated: March 19, 2026