""" z_ev_specs.py — pure gosling spec builders for z_external_validation_fork.py. Sibling companion to `z_gb_specs.py`. Same contract: depends only on the `gosling` API object passed in (no pandas, no marimo), so every builder can be schema-validated headlessly (`gos..validate()`). The notebook does all the pandas shaping and hands these builders plain lists-of-dicts + the imported `gos` module + a colour dict `C`. The external-validation fork's three cross-checks each have their OWN coordinate-bearing evidence table, so each gets a faithful *genomic* view (the parent notebook only ever showed them as summary bars / a confusion matrix — this adds the "where on the genome" layer, the way the genome-browser fork did for the arcs): build_crispr the EPCrisprBenchmark element→gene pairs as whole-genome ARCS, coloured by confusion-matrix category (TP / FP / FN / TN) with a CRISPRi-effect-size SKYLINE — so the ~37× odds ratio becomes a genomic picture of *which* perturbation-validated pairs our map recovers vs misses. Queryable by CRISPR cell type. The eQTL genetic-support view and the ρ link-strength / metadata explorer REUSE the genome-browser fork's already-perfected builders (`z_gb_specs.build_locus` for the fine-mapped-variant lollipops, `z_gb_specs.build_program` for the whole-genome ρ arc-skyline), so the two forks render identically and there is one place to maintain each idiom. This module only adds what the browser fork lacked: the CRISPR confusion-category genomic view. Rendering facts (identical to z_gb_specs — kept in sync): 1. At whole-genome scale a cis link is sub-pixel, so an arc track is paired with a point SKYLINE that stays legible at any zoom; the arcs resolve as you zoom into an arm (shared linkingId). 2. A thin `withinLink` shows only its STROKE, and an unset stroke defaults to black — so links are coloured on `stroke`, not `color`. 3. A `type:nominal` colour channel takes a custom `domain`/`range`; a `type:quantitative` one is given gosling's own viridis ramp (custom ranges are ignored) — so categories get explicit colours, ρ keeps the shared viridis domain from z_gb_specs. """ from __future__ import annotations # reuse the browser fork's shared genome constants / helpers so the two forks stay pixel-identical # and there is a single source of truth for the ideogram, chromosome bands and ρ colour scale. from z_gb_specs import ( HG38, CHR_ORDER, RHO_DOMAIN, _present, _ideo_data, _chr_bands, _rho_color, build_locus, build_program, build_atlas, ) # the four confusion-matrix categories, ordered recovered→missed, with the notebook's palette roles. # (Kept here, not in C, so the legend order is stable regardless of dict iteration order.) CRISPR_CATS = ["TP", "FP", "FN", "TN"] CRISPR_CAT_LONG = { "TP": "TP · Regulated & we link", "FP": "FP · not Regulated but we link", "FN": "FN · Regulated, we miss", "TN": "TN · not Regulated & we don't link", } def _cat_color(gos, C, legend): """The one confusion-category colour scale, shared by the CRISPR arcs and effect-size skyline so a single legend reads for both. GREEN=TP, RED=FP, GREY=FN, SLATE=TN — matching §2's matrix.""" long = [CRISPR_CAT_LONG[c] for c in CRISPR_CATS] rng = [C["GREEN"], C["RED"], C["GREY"], C["SLATE"]] return gos.Color("cat_long:N", domain=long, range=rng, legend=legend) # ------------------------------------------------------------------------------------------------ # CRISPR — the EPCrisprBenchmark pairs, genomic (arcs by confusion category · effect-size skyline) # ------------------------------------------------------------------------------------------------ def build_crispr(gos, C, *, pairs, y_domain, width=860): """ pairs : [{chrom, e_mid, tss, cat, cat_long, effect}] one CRISPR-tested element→gene pair each. cat ∈ {TP,FP,FN,TN}; cat_long is the legend label; effect is the CRISPRi EffectSize. y_domain : [lo, hi] the observed EffectSize spread, pinned so the skyline resolves it. Whole-genome, LINKED (shared linkingId → ideogram + arcs + skyline pan/zoom together): · chromosome ideogram · a real ARC track (element→gene), coloured by confusion category on the STROKE (fact #2) · an effect-size SKYLINE (one point per pair, y = CRISPRi EffectSize), same category colours — the always-legible overview; the arcs (sub-pixel zoomed out, fact #1) resolve as you zoom in. Answers, genomically: *where do the perturbation-validated pairs sit, and which does our map recover (TP) vs miss (FN) vs fire on wrongly (FP)?* """ present = _present(a["chrom"] for a in pairs) LID = "crisprwg" ideo = (gos.Track(_ideo_data(gos, present)).mark_rect() .encode(x=gos.X("s:G", axis="top", linkingId=LID), xe="e:G", color=_chr_bands(gos, present)) .properties(width=width, height=12)) data = gos.Data(type="json", values=pairs, chromosomeField="chrom", genomicFields=["e_mid", "tss"]) arcs = (gos.Track(data).mark_withinLink() .encode(x=gos.X("e_mid:G", axis="none", linkingId=LID), xe="tss:G", stroke=_cat_color(gos, C, False), strokeWidth=gos.value(1.1), opacity=gos.value(0.55)) .properties(width=width, height=120)) sky = (gos.Track(data).mark_point() .encode(x=gos.X("e_mid:G", axis="none", linkingId=LID), y=gos.Y("effect:Q", domain=list(y_domain), axis="left"), color=_cat_color(gos, C, True), size=gos.value(5.0), opacity=gos.value(0.8), stroke=gos.value("white"), strokeWidth=gos.value(0.3)) .properties(width=width, height=190)) return gos.vertical(ideo, arcs, sky).properties(spacing=4).to_dict() # thin, self-documenting re-exports so the notebook imports one module. These ARE the browser fork's # builders — the eQTL lollipop locus and the ρ arc-skyline are already perfected there; the external- # validation fork feeds them its own (eQTL-supported / metadata-filtered) rows. build_eqtl_locus = build_locus # cCRE rects + fine-mapped-variant lollipops on a PIP axis build_links = build_program # whole-genome ρ arc + skyline, or circular ρ chords build_links_atlas = build_atlas # every arc, circular, coloured by chromosome