Do Neftel-style GBM state calls replicate across platforms? A GlioState re-analysis (draft)
Review pending Version 0.1.0 · status: draft · last expert review: not yet reviewed
Research question
- Do GBM cell-state calls replicate on an independent platform, and does the state structure go beyond the signature genes?
- How much does a trivial baseline recover?
Current understanding
In a GlioState re-analysis of public data (GSE131928), tumor-level state composition replicated moderately across Smart-seq2 and 10X (Spearman ρ=0.82 over 24 tumor×state values in 6 shared tumors; dominant state matched in 4 of 6), and a classifier using no signature genes recovered signature-based state calls on held-out 10X cells with balanced accuracy 0.75 (chance 0.25; single-marker baseline 0.53). This tests the reproducibility of state definitions only; it is not a test of any therapy-response or virtual-cell model, and the labels are not ground truth.
Why this question: the library contains no GBM-specific benchmark of state-prediction models, and perturbation-prediction benchmarks in other systems report that deep models do not yet reliably beat simple baselines. A virtual cell for GBM would first need state labels that are reproducible across datasets and platforms. VC0048 VC0049 VC0050 VC0059
Result in brief: state calls based on the Neftel meta-modules carry transcriptome-wide structure that transfers across platforms, clearly above chance and above a single-marker baseline, but with substantial tumor-to-tumor variability (per-tumor balanced accuracy 0.50–0.86) and notable confusion of OPC-like and AC-like cells with MES-like cells. VC0059
Evidence map
Primary studies & datasets
- VC0059 An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. — Neftel C et al., Cell, 2019
- VC0060 Charting the single-cell and spatial landscape of IDH-wild-type glioblastoma with GBmap. — Ruiz-Moreno C et al., Neuro-oncology, 2025
- VC0069 A single-cell atlas of glioblastoma evolution under therapy reveals cell-intrinsic and cell-extrinsic therapeutic targets. — Wang L, Jung J, Babikir H, Shamardani K, Jain S, Feng X, Gupta N, Rosi S, Chang S, Raleigh D, Solomon D, Phillips JJ, Diaz AA., Nature cancer, 2022
- VC0092 · preprint Mapping the spatial architecture of glioblastoma from core to edge delineates niche-specific tumor cell states and intercellular interactions — Khan SM et al., bioRxiv, 2025
Related models & tools
- VC0048 Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. — Ahlmann-Eltze C, Huber W, Anders S., Nature methods, 2025
- VC0049 Benchmarking foundation cell models for post-perturbation RNA-seq prediction. — Csendes G, Sanz G, Szalay KZ, Szalai B., BMC genomics, 2025
- VC0050 Zero-shot evaluation reveals limitations of single-cell foundation models. — Kedzierska KZ, Crawford L, Amini AP, Lu AX., Genome biology, 2025
Perturbation & validation evidence
- VC0070 In vivo perturb-seq of cancer and microenvironment cells dissects oncologic drivers and radiotherapy responses in glioblastoma. — Liu SJ et al., Genome biology, 2024
8 library items cited on this page; 1 are preprints (not peer reviewed).
Relations
Typed relations between entities, each backed by library items (role = how the item bears on the relation). Low-confidence rows need editorial review. All entities →
| Subject | Relation | Object | Confidence | Evidence (role) |
|---|---|---|---|---|
| Neftel meta-modules | marks | Mesenchymal-like state | medium | VC0059 supports Signature-score calls from the Neftel meta-modules are not ground truth; the signatures used in GlioState’s re-analysis were an unverified third-party transcription. |
| Neftel meta-modules | marks | Neural-progenitor-like state | medium | VC0059 supports Signature-score calls from the Neftel meta-modules are not ground truth; the signatures used in GlioState’s re-analysis were an unverified third-party transcription. |
| Neftel meta-modules | marks | Oligodendrocyte-progenitor-like state | medium | VC0059 supports Signature-score calls from the Neftel meta-modules are not ground truth; the signatures used in GlioState’s re-analysis were an unverified third-party transcription. |
| Neftel meta-modules | marks | Astrocyte-like state | medium | VC0059 supports Signature-score calls from the Neftel meta-modules are not ground truth; the signatures used in GlioState’s re-analysis were an unverified third-party transcription. |
| Neftel four-state framework | uses dataset | GSE131928 | medium | VC0059 dataset Accession as stated in the GlioState re-analysis text (Neftel et al. data on GEO). |
Entities on this page
- Astrocyte-like state Cell states
- Glioblastoma Diseases
- GSE131928 Datasets
- Mesenchymal-like state Cell states
- Neftel four-state framework Methods & concepts
- Neftel meta-modules Methods & concepts
- Oligodendrocyte-progenitor-like state Cell states
Methods and results (our own analysis)
This is an exploratory re-analysis of public data run by GlioState, not a peer-reviewed result. Outputs: results/results.json · results/E1_shared_tumor_composition.csv · README.md (the analysis script itself is not published on this site)
Data and licence
Data: GEO GSE131928 (Neftel et al.; processed TPM only): Smart-seq2 7,930 cells (GSM3828672, ~311 MB gz) and 10X 16,201 cells (GSM3828673, ~358 MB gz) from 28 tumors; 6 tumors (MGH102, 105, 115, 124, 125, 143) were profiled on both platforms. Public GEO data; reuse terms were not independently checked. Signatures: a third-party GitHub transcription of the Neftel meta-modules (no licence file; not checked against the paper’s supplementary table). Marker genes quoted in the paper text were confirmed present in the expected modules; the AC module has 39 genes. VC0059
Design (fixed before viewing results)
Cells were scored with a Tirosh-style expression-matched control-subtracted mean for four collapsed states (MES, AC, OPC, NPC) and assigned to the highest score. Immune-marker cells (PTPRC, CD14 or CD68 >2 on the log2(TPM/10+1) scale) were removed; no copy-number-based malignant-cell calling was done. E1: composition concordance between platforms in the 6 shared tumors. E2: a logistic-regression classifier trained on Smart-seq2 cells from the 22 tumors not profiled on 10X, using the 1,500 most variable genes excluding all signature and marker genes, tested on all 7,930 retained 10X cells (external platform; platforms standardised separately, no test labels used). Baselines: majority class, single marker per state (GFAP, OLIG1, DCX, VIM), label-permutation null (100 permutations); uncertainty by tumor-cluster bootstrap and leave-one-tumor-out.
Results
| Metric | Value |
|---|---|
| E1 composition concordance (6 shared tumors, 24 values) | Spearman ρ = 0.82; tumor-shuffle permutation p = 0.004; dominant state matched in 4/6 tumors |
| E2 balanced accuracy, signature-free classifier → 10X | 0.746 (tumor-cluster bootstrap 95% CI 0.67–0.84); macro-F1 0.69; Cohen κ 0.59 |
| Majority-class baseline | 0.25 (κ 0) |
| Single-marker baseline (GFAP/OLIG1/DCX/VIM) | 0.53 (macro-F1 0.46; κ 0.27) |
| Label-permutation null | mean 0.24; 95th percentile 0.27 |
| Leave-one-tumor-out within Smart-seq2 (22 tumors) | median 0.69 (IQR 0.58–0.73) |
| Per-tumor balanced accuracy on 10X (9 tumors) | range 0.50 (MGH126) – 0.86 (MGH115) |
| Cells with small score margin (<0.1; n=393) | balanced accuracy 0.40 vs 0.76 for confident cells |
Effect sizes should be read with the failures: in MGH105 the MES-like fraction was 26% on Smart-seq2 but 72% on 10X; in MGH125 the NPC-like fraction was 66% vs 35%; MGH124 OPC-like 29% vs 7%. The classifier labelled 632 of 2,667 reference AC-like cells and 591 of 1,805 reference OPC-like cells as MES-like. Overall MES-like fraction was higher on 10X (36%) than on Smart-seq2 (16%).
A first run contained an error (10X data not log-transformed, immune filter too aggressive); it was corrected, and the flawed first-run output is not published on this site.
What remains uncertain
Labels are signature-score calls, not ground truth; E2 shows structure beyond the signature genes but is partly circular because the same labels train and score the classifier. Independent genotype or lineage-tracing validation was not available. VC0059
Signatures are an unverified third-party transcription; the official supplementary table was not used.
Malignant cells were not identified by copy-number inference; the crude marker filter removed 51% of 10X cells and 20% of Smart-seq2 cells, and residual non-malignant astrocyte/oligodendrocyte-lineage cells may inflate AC/OPC/MES calls. The platform difference in MES-like fraction may reflect technical effects (e.g. stress response, ambient RNA) rather than biology; this was not tested.
Only 28 tumors, 6 shared across platforms, mostly from one centre; results from GBmap (an independent 26-dataset atlas) or other cohorts were not tested here. Spatial and therapy-associated states were not assessed. VC0060 VC0069
This analysis does not test any perturbation or therapy-response model, and does not evaluate whether state transitions under treatment can be predicted. VC0070 VC0092 · preprint
For research and education only. Not medical advice and not a recommendation of any diagnosis, treatment or product. Patients should consult their own physicians.