Atlas — per-cohort ThermoCas9 target shortlists

Headline figure — per-positive whole-genome rank percentiles

Figure 4. For each of the three Roth Fig. 5d positives (ESR1, EGFLAM, GATA3), the dot-plot shows the rank percentile of that positive's PAM site within the cohort's whole-genome candidate universe (HM450 ≈ 19.8 M loci; EPIC v2 ≈ 35.4 M). Filled circles are V2.5-sigmoid; open circles are V2.5-diff; the line connects the two so the rank-lift direction is immediately visible. Underlying data: per_positive_wg_percentile.json · SVG.

What to read off the plot.

• On the three matched cell-line cohorts (left three columns), V2.5-diff and V2.5-sigmoid sit on top of each other above the 95-percentile line — AUC parity within 0.002. The tie_band@100 collapse from 421–1,493 records under V2.5-diff to 1 under V2.5-sigmoid is the actual usability story on these cohorts; this dot-plot only shows that the ranks are equivalent in position.
• On the GSE69914 tissue cohort (right column), the dumbbells stretch — and not all in the same direction. GATA3 and EGFLAM improve substantially under V2.5-sigmoid; ESR1 moves the wrong way. This is the non-uniform-superiority caveat the methods paper discloses in §6.1: in the GSE69914 EXACT + PROXIMAL_CLOSE restricted-universe subset (where ESR1 is the only evaluable positive), V2.5-sigmoid trails V2.5-diff, raw Δβ-only, and the limma-style baseline.

Per-cohort top-100 explorer

Pick a cohort, filter by gene symbol or PAM family. Rows highlighted in the accent color overlap a Roth Fig. 5d positive at the wide (±500 bp) tier. The full per-row schema (30+ columns including RepeatMasker overlap and CGI distance) lives in the per-cohort TSV linked below; this table shows the slim-schema columns that fit on a screen.

Summary by cohort

How rows on this page map to rows in the paper

• The dot-plot (fig4) visualizes the per-positive WG-rank table reported numerically in PAPER §5.2.2. The data file (examples/genome_wide_panel.md) is committed at the paper tag.
• The per-cohort top-100 tables are the same artifact PAPER §5.5 reports at top-20, expanded to top-100 for this website. They ship as examples/<cohort>_roth_labels/top100_atlas.{tsv,md} at memo-2026-04-22-bw.
• The tie-band collapse story (V2.5-diff 421–1,493 → V2.5-sigmoid 1) is visible as the y-axis collapse in PAPER fig2; this atlas page does not duplicate that figure.
• The ESR1 reversal is highlighted graphically in the dot-plot above and discussed quantitatively in PAPER §5.7.

What is not in this atlas (intentionally)

• No TCGA pan-cancer rows. The framework includes a streaming k-way-merge pan-cancer aggregator and unpublished TCGA cohort scoring exists in the working tree, but those rows are not on this surface yet because they have not been audited against a tagged paper artifact and because publishing pan-cancer top-K shortlists on a corporate domain reads as a drug-discovery pipeline more than a benchmarked methods demonstration. A future v2 atlas may add a TCGA section behind explicit framing.
• No genome-browser / Manhattan view. The UCSC Genome Browser already does this well; click any candidate_id chromosome coordinate from the per-cohort TSVs to jump there.
• No score-distribution histograms. Those live under examples/*_roth_labels/ as supplementary materials; this page intentionally scopes to "where do the candidates land" rather than "what does the score distribution look like."

Cite or reproduce

• Methods paper — PAPER.pdf at tag paper-5-10j; Bioinformatics-shaped short version MANUSCRIPT.pdf at tag memo-2026-04-22-bw.
• Atlas reproducer scripts — build_atlas_dotplot.py and build_atlas_top100.py.
• Roth et al., Nature 2026 — read the article on nature.com (DOI 10.1038/s41586-026-10384-z).