Glioblastoma invasion gets a live-brain-slice tracker: thousands of cells followed as standard therapy enriches the fastest branching state
Synopsis
Glioblastoma cells switch among branching, diffuse translocation, locomotion and perivascular invasion modes, but these switches were hard to observe at scale over days in intact brain tissue. GlioTrace links whole-specimen confocal imaging, convolutional-network morphology classification and a covariate hidden Markov model to track thousands of cells across six patient-derived xenograft brain slices. Invasion plasticity proved patient-specific and spatially organized, modulated by vascular proximity and microglial contact; targeted drugs shifted phenotype composition and transition probabilities, while radiation plus temozolomide enriched the fastest, highly connected branching state.
Interpretation
Glioblastoma invasion-state switching can now be measured continuously in intact brain tissue rather than inferred from cells fixed at a single time point. Earlier organotypic cultures seeded exogenous tumor cells onto pre-cut naive brain slices and did not recapitulate the architecture established during orthotopic growth; cranial-window intravital imaging reached single-cell resolution but was restricted to superficial cortical fields and could not provide the full cross-sectional sampling needed for population-level phenotyping. Across six patient-derived xenograft models from the HGCC collection, imaging covered 300-micron brain slices at 10X resolution over 4 to 6 days, with roughly 100 time points at intervals under 2 hours, producing 9000 by 9000 pixel videos able to resolve slender cellular protrusions.
Each cell is assigned at each time point to one of five morphological classes, and switches among those classes are reconstructed with a probabilistic model. Frame-wise morphology classification carries no temporal consistency and is noisy, so a state model is needed to connect evidence across adjacent frames. The convolutional network was trained on 7800 manually curated images to separate branching, diffuse translocation, locomotion, perivascular translocation and round classes, with an extra crowded or outlier class excluded from downstream analysis, reaching 87% overall accuracy; the hidden Markov model uses the network posteriors as observations and vascular distance, microglial colocalization and local polarization as covariates on transition probabilities.
Therapeutic perturbation changes invasion phenotype composition and the state-transition network, not only cell speed. Against matched untreated slices, dasatinib and thapsigargin both reduced cell speed in most models and raised the round-cell fraction while reducing invasive morphologies such as branching and locomotion. Dasatinib and thapsigargin both eliminated branching cells in U3013MG; U3220MG retained branching cells after dasatinib but lost them after thapsigargin. After 20 Gy in vivo radiation plus temozolomide, the branching fraction rose consistently in some models and transition networks showed reinforced self-transitions, meaning cells more often stayed in their current state.
Perspective
The demonstrated scope is ex vivo imaging of six HGCC patient-derived xenograft models in 300-micron brain slices, spanning proneural, mesenchymal and classical transcriptional subtypes and retaining intact vasculature and resident microglia; the pipeline runs on standard high-content confocal microscopes and the accompanying analysis software is released as an open-source Python package.
The models are established in immunodeficient mice lacking functional T and B lymphocytes, so adaptive immune interactions are absent and microglial activation states may differ from immunocompetent tumors; slicing introduces surface injury and the air-liquid interface does not recapitulate the blood-brain barrier, which matters when interpreting responses to agents whose efficacy depends on CNS penetration; the current implementation is restricted to two fluorescence channels and maximum intensity projections rather than full three-dimensional volumes, and six models are insufficient for statistically powered subtype-level conclusions.