A harmonized human glioblastoma chromatin accessibility atlas (snATAC-seq, 48 patients, 8 cohorts, 1.52M nuclei)

Harmonized snATAC-seq / snMultiome atlas across 8 published GBM cohorts (48 patients with ≥50 cells per label group, 1,523,668 nuclei) harmonized to a 1,958,753-peak union reference (74.2% de-novo vs CATLAS), with per-cell-type specificity and strength scoring designed for AAV-delivered cell-type-specific enhancer therapy nomination. Primary ranking covers 10 primary cell types (astrocyte, microglia, oligodendrocyte, OPC, endothelial, neuron, GABA_neuron, T_cell, TAM, malignant_pooled) and a 13-way taxonomy that splits malignant into 4 Neftel states (AC / MES / NPC / OPC). Scoring (v4, built 2026-10-17): primary_score = mean of two within-cell-type percentile ranks — spec_pct (percentile rank of log2FC_top_vs_next) and strength_pct (percentile rank of patient-averaged, width-normalized Tn5 fragments per cell of target type, with "evaluable" = patient has ≥100 cells of target type). No model assumption, no Z-scores, no copy-number correction. CNV is handled diagnostically only: non-malignant peaks within ±1 Mb of chr7p11.2 EGFR, chr12q15 MDM2, chr8q24 MYC, or chr9p21.3 CDKN2A (deletion) carry an AMP badge flagging ambient-fragment contamination from the dominant malignant population (Spearman ρ = +0.70 between non-malignant chr7p signal and per-patient malignant-cell fraction, p < 0.001). Replication is reported alongside the ranking (patient_detection_rate, n_cohorts) but is not folded into primary_score. Peaks within 2 kb of the nearest annotated TSS are excluded from ranking (95.9% survive). Per-cell QC drops barcodes with nnz < 1,000 union-peak events before any aggregation. Canonical cohort slugs (first-author_year): greenwald_2024 (Greenwald Cell / E-MTAB-17183), sojka_2025 (Sojka Nat Cell Biol / GSE276177 — the earlier slug 'khan_astro' was a misattribution, corrected in v2026-10-11), sundaram_2024 (Sundaram Cell Rep / GDC a7765b5f, open-access TCGA scATAC subset), zhao_2024 (Zhao CRM / HRA004942), mathewson_2021 (Mathewson Cell / GSE163654 + Wang 2024 Sci Adv), guo_2020 (Guo Nat Commun / GSE138794), guilhamon_2021 (Guilhamon eLife / GSE139136), raviram_2023 (Raviram et al. 2023 PNAS 120:e2210991120, sci-ATAC / GSE165037). All 8 cohorts contribute to BOTH the peak reference and scoring. Deposit layout: a single 2.0 GB bundle GBM-atlas-v2026-11-02.tar.gz contains the full atlas tree (reports/, data/, scripts/, tests/, workflow/, supplementary/, reference/, metadata, CITATION.cff, SCHEMA.md, MANIFEST.sha256). Five files are also exposed at the Zenodo landing for quick inspection without untarring: README.md, CHANGELOG.md, atlas_v4.html (27 MB self-contained interactive viewer with 14 cell-type tabs, sortable + filterable), atlas_v4_SCORING_README.md (methodology), and atlas_v4_viewer_bundle_README.txt. SHA256SUMS.txt holds integrity hashes for the above landing files and the tarball. Inside the bundle: reports/atlas_v4/ holds the primary ranking (atlas_v4.html, per_cell_type_top1000/ TSVs, primary_score_summary.tsv, v4_master_scores.parquet with all 1.96M peaks × 14 cell types, atlas_v4_slim.csv, atlas_v4_supplementary_wide.parquet). reports/atlas_selective_strength/ holds a strength-first selective cross-check (CPM ≥ 10, log2FC ≥ 2). reports/Other analysis/ preserves the pre-v4 top-100 polished reports from the 2026-10-14 release for legacy citation. data/matrix/ holds enhancer_candidate_matrix.parquet (1.96M peaks × 337 patient×cell-type pseudobulks, CPM-normalized), greenwald_fine_pseudobulk_backup.parquet (99-label fine-grained Greenwald pseudobulk preserved for future re-remapping), union_peak_reference.bed, peak_provenance.parquet, and slug_canonical.tsv. metadata_2026-10-27.csv + metadata_2026-10-27_fullcols.csv provide 48-donor sample metadata; cohort_attribution.csv provides the 8-cohort provenance table. Pipeline source: github.com/jpmcginnis/HumanAAVProject (Glioblastoma snATACseq pipeline subdirectory).