Analysis code for Decoding Plasticity Regulators and Transition Trajectories in Glioblastoma with Single-cell Multiomics — Saraswat M, Rueda-Gensini L, Heinzelmann E, Gracia T, Memi T et al.
If you use code from this repository, please cite our paper Much of the analysis here is downstream of scDoRI (bioFAM/scDoRI). See Citation.
Regulatory output and supporting data are archived on Zenodo: https://doi.org/10.5281/zenodo.21873381 see Data availability.
Method implementation: https://github.com/bioFAM/scDoRI (archived at https://doi.org/10.5281/zenodo.21937078)
| data it reads | ||
|---|---|---|
TAP_TRS_analysis.ipynb |
Topic activation potential (TAP), topic repression score (TRS), net topic transitions, contribution of epigenetic priming, state-level regulation | data/ only |
plot_supp_peak_object.ipynb |
Peak-level panels: peak-topic UMAPs, accessibility per state, TF binding, gene–peak links | Zenodo peak AnnData |
plasticity_analysis.ipynb |
Epigenetic plasticity — entropy of ATAC-based cell-state classifier probabilities | ArrayExpress metacell objects |
python_scripts/topic_regulation.py |
Module behind the TAP/TRS notebook: TF activity scoring, topic- and state-level regulation potentials, target-topic scaling, randomised backgrounds | — |
tf_screen/ |
55-TF overexpression screen pipeline, Fig. 3 and ED Fig. 5 — own README | GEO |
Only TAP_TRS_analysis.ipynb runs from the repository alone. Everything it needs is
in data/. The other two notebooks read objects that are too large for git and must be
downloaded first — the peak AnnData from Zenodo, the metacell objects from ArrayExpress.
Python 3.10+ with numpy, pandas, scipy, scikit-learn, matplotlib, seaborn and
tqdm. TAP_TRS_analysis.ipynb needs nothing beyond these.
Use pandas < 3. python_scripts/topic_regulation.py relies on chained assignment
(df[col][mask] = 0), which pandas 3 turns into a silent no-op under copy-on-write; the
act_thresh filter then stops taking effect and the state-level cells fail. The published
results were produced with pandas 2.x.
All files below are in data/ and are read directly by TAP_TRS_analysis.ipynb.
| file | shape | role |
|---|---|---|
data/01_scdori_grn_activator_atac.npz |
50 topics × 195 TFs × 3,192 genes | activating TF→gene links per topic (ATAC-based) |
data/01_scdori_grn_repressor_finetuned.npz |
50 topics × 195 TFs × 3,192 genes | repressing TF→gene links per topic (fine-tuned) |
data/02_scdori_topic_tf_activity_activator.csv |
50 topics × 195 TFs | activator TF activity per topic |
data/02_scdori_tf_expression_by_topic.csv |
195 TFs × 20 topics | mean scaled TF expression per topic |
data/02_scdori_tf_geneactivity_by_topic.csv |
195 TFs × 20 topics | mean scaled TF gene activity score (ATAC based accesibility) per topic |
data/02_scdori_topic_gene_matrix_selected31.csv |
3,192 genes × 31 topics | topic-gene loadings; also supplies the gene order of the GRN tensors |
data/02_scdori_mean_topics_activation_coarse_states.tsv |
5 states × 50 topics | topic activity per coarse state, for the state-level analysis |
data/04_TAP_random_background.csv |
342 topic pairs × 1,000 permutations | null distribution of TAP values |
data/04_TRS_random_background.csv |
342 topic pairs × 1,000 permutations | null distribution of TRS values |
Each .npz stores its own topics, tfs and genes label arrays next to grn
The permutation nulls are shipped precomputed, so significance masking runs in seconds — the
notebook loads the two 04_*_random_background.csv files and zeroes any topic pair whose
value falls below the 95th percentile of its own null. Recomputing the nulls from scratch
(compute_significance=True, 1,000 randomisations per topic pair) takes hours and is not
needed to reproduce the figures.
The 342 rows are the ordered pairs of 19 topics — the 17 analysed plus Topics 31 and 39 — with self-pairs excluded. The notebook subsets them to the 17 topics it plots. Row order is significant: the notebook reshapes the column of thresholds into a 19 × 18 matrix, so the rows must stay in the file's original order.
| Archive | Accession | Contents |
|---|---|---|
| Web portal | www.gbmspace.org | Interactive browser for the processed multiome atlas |
| ArrayExpress | E-MTAB-17183 |
Processed snRNA-seq and snATAC-seq AnnData; metacell-level RNA counts and ATAC gene scores (plasticity analysis) |
| EGA | EGAD00001015526 |
Raw multiome sequencing data (controlled access) |
| GEO | GSE294518 |
All remaining NGS data, including the TF screen |
| Zenodo | 10.5281/zenodo.21873381 |
scDoRI regulatory output + SCENIC+ comparison networks |
Reference genome: hg38 (GRCh38).
Zenodo is the citable archive. It holds the same TAP/TRS inputs that are committed here,
plus everything too large for git, and has its own 00_README.md:
- Three scDoRI GRN tensors (ATAC-based activator, fine-tuned activator, fine-tuned
repressor), each 50 topics × 195 TFs × 3,192 genes, as
.npz— plus the same links as one edge-list table. Only the first and third are needed here; the fine-tuned activator is Zenodo-only - Topic-level matrices: topic-gene loadings, TF activity per topic, TF expression and gene activity per topic, topic activity per coarse state, topic annotation
- SCENIC+ networks (global + nine per cell type) and per-topic AUCell enrichment
- Peak AnnData: 182,677 peaks × 195 TFs — in-silico ChIP binding, peak-topic activity, accessibility per state, peak UMAP, gene–peak links
Two activator GRNs, not interchangeable. The ATAC-based one is not fine-tuned against TF–gene expression and is what TAP uses; the fine-tuned one is anchored to observed expression.
Repressor weights are negative. In both the GRN tensors and the peak object. Percentile colour limits applied naively will highlight the weakest sites.
Raw TF binding is not comparable between TFs. Standardise with the per-TF mean and sd stored in the peak object before comparing factors.
scale_topic_regulation_target_topic modifies its input in place. It divides each row
by the number of TFs expressed in that source topic before scaling columns, so calling it
twice on the same object double-normalises. The notebook recomputes the raw matrix before
each call for this reason.
<FULL PAPER CITATION Upcoming, Nature 2026>
MIT — see LICENSE. Zenodo data is CC BY 4.0.
Open an issue, or contact manu.saraswat AT dkfz.de or laura.ruedagensini AT dkfz.de