v3 primary ranking (rank-percentile sum): Top-1000 per cell type ranked by
v3 strength (fragments-per-cell proxy):
Distal filter: |dist_to_tss| ≥ 2 kb applied before top-N selection. Mirrors GBM v4 convention — promoter-proximal peaks aren't useful AAV enhancer scaffolds and would otherwise dominate the top ranks.
Peak set is v1.0.0 (pre-Design-B). This report is scored against the 990K pre-Design-B union peak set, median width 895 bp (mean 2,434 bp), which carries wider peaks than the current Design-B narrow union (2.06M peaks, median 518 bp). Chrom/start/end coordinates are correct; coord lookups in UCSC work. For AAV cloning, use the sister script
Sparse-cell-call caveat. The 4DN hippocampus cohort (25.7M barcodes) and GEO GSE322785 (9.7M barcodes) shipped with permissive upstream barcode-calling; median nnz per barcode is 3-4 in those cohorts vs 5,000-15,000 in the rest. The upstream v2.0 pipeline inherited whatever filtering was applied by the author deposits. If a peak's apparent cell-type signal is dominated by patient counts from 4DN or GSE322785 and no other cohort replicates, treat the specificity call with caution; cross-check
Nearest-gene is informational only. AAV scaffold efficacy depends on cell-type-selective driving strength (what primary_score ranks), not nearest-gene identity. Median mammalian enhancer-promoter distance is ~50-100 kb with a long tail beyond 1 Mb.
Hover any column header for its definition and formula.
primary_score = (spec_pct + strength_pct) / 2, where spec_pct is the percentile rank of
log2FC_top_vs_next = log2(mean_frac_target / off_target_max) and strength_pct is the percentile rank of
mean_frac, both computed within distal peaks of each cell type.
v3 strength (fragments-per-cell proxy):
strength = mean across evaluable donors of fraction_open, where fraction_open = cells accessing peak / total cells of that class per donor.
Same semantic as GBM v4's strength_frags_per_cell_wn but computed from v2.0 patient-level pseudobulks (not re-quantified at a shared peak width — see peak-set caveat below).
Distal filter: |dist_to_tss| ≥ 2 kb applied before top-N selection. Mirrors GBM v4 convention — promoter-proximal peaks aren't useful AAV enhancer scaffolds and would otherwise dominate the top ranks.
Peak set is v1.0.0 (pre-Design-B). This report is scored against the 990K pre-Design-B union peak set, median width 895 bp (mean 2,434 bp), which carries wider peaks than the current Design-B narrow union (2.06M peaks, median 518 bp). Chrom/start/end coordinates are correct; coord lookups in UCSC work. For AAV cloning, use the sister script
narrow_lookup.py to resolve a top-ranked wide peak to the narrow Design B sub-peak(s) with the actual cloneable sequence.
Sparse-cell-call caveat. The 4DN hippocampus cohort (25.7M barcodes) and GEO GSE322785 (9.7M barcodes) shipped with permissive upstream barcode-calling; median nnz per barcode is 3-4 in those cohorts vs 5,000-15,000 in the rest. The upstream v2.0 pipeline inherited whatever filtering was applied by the author deposits. If a peak's apparent cell-type signal is dominated by patient counts from 4DN or GSE322785 and no other cohort replicates, treat the specificity call with caution; cross-check
n_cohorts (should be ≥ 3) and surgical/postmortem agreement.
Nearest-gene is informational only. AAV scaffold efficacy depends on cell-type-selective driving strength (what primary_score ranks), not nearest-gene identity. Median mammalian enhancer-promoter distance is ~50-100 kb with a long tail beyond 1 Mb.
Hover any column header for its definition and formula.
| peak_id | chrom | start | end | width (bp) | nearest gene | dist to TSS | band | log2FC top vs next | spec pct | strength (mean frac) | strength pct | primary score | n cohorts | donor detection |
|---|