What this notebook asks. Some pigmentation-associated loci sit next to a gene that is not a known pigment gene, so which gene actually drives the trait is unresolved. This notebook tests whether each such locus can be rescued — linked into the melanin-making (melanogenesis) network by an independent, non-circular line of evidence — and grades how strong each link is. Loci with no such link are reported as negatives alongside the positives.
Scope note (reconciles with NB4). These 18 come from a programmatic pass over Ang2023 + Morgan2018. The claim-faithful re-classification in NB4 refines the effector-uncertain target to 14 within the 105 curated loci (34 with Kim 2024 folded in). Of these 18, that re-read keeps 9 as still-uncertain, 6 as ambiguous-near, and reclassifies 3 as canonical/regulatory — consistent with this diagnostic’s own finding that the “uncertain” label was often wrong (e.g. LRMDA below).
Of these 18 provisionally effector-uncertain loci (nearest gene is NOT a canonical pigmentation effector), asking whether the candidate gene reaches melanogenesis via a non-circular route (direct substrate edge, MITF/SOX10/PAX3 GRN regulon, Reactome pigmentation subpathway [gated on the pigmentation-vs-MITF-program tag], melanocyte eQTL [Zhang 2018, tissue-correct], ≤2-hop shortest path, or verifiable known biology):
| bucket | n | loci |
|---|---|---|
| connected via NON-canonical route | 5 | LRMDA, EGFR, ATRN, SIK1, CCND1 |
| connected only via canonical neighbor (NOT our finding) | 5 | IRF4, TYR(GRM5), SLC24A4, MC1R(FANCA), POMC(EFR3B) |
| needs manual paper review | 2 | MSX2, PKHD1 |
| no connection (negative) | 6 | SYT6, KALRN, TECRL, CNKSR3, RAB11FIP2, PPP3CA-segment |
But NONE of the 5 is a novel, network-discovered effector. Ranked: 1. LRMDA (strongest, grade A) — is literally OCA7 albinism gene (“leucine-rich melanocyte differentiation associated”), a strong melanocyte eGene (q=3.9e-11, as alias C10orf11). This is not a discovery — the effector-uncertain label was wrong; it should never have been in the uncertain set. 2. SIK1 (grade C, solid) — textbook MC1R→cAMP→PKA⊣SIK1→CRTC→CREB→MITF axis; GWAS p=1e-8, replicated. Real biology, but “known,” not network-derived. 3. ATRN (grade C) — attractin = mouse mahogany coat-color modifier (agouti/ASIP–MC1R). Real biology, but Ang2023 GWAS is null (p_adj=0.14). 4. EGFR (grade C) — genuine MITF-repressed / PAX3-activated GRN target (GRN layer uniquely enables). But Ang2023 GWAS null (p_adj=0.78, not replicated). 5. CCND1 (weakest) — MITF target, but a proliferation output (borderline circular) and bulk-only (NOT a melanocyte eGene); melanocyte data does not support the resolution.
Bottom line: almost nothing NEW connects. The rescue screen surfaces no novel melanogenesis effector. Its “hits” are (a) an already-known albinism gene the extraction mislabeled, (b) textbook biology at loci whose GWAS signal is often weak/null, or (c) generic canonical-neighbor regulatory signals. The 6 negatives + 2 manual-review loci have no melanogenesis route at all. Melanocyte-vs-bulk correction matters: it retracts three bulk L2G calls (TECRL, RAB11FIP2, CCND1) and confirms LRMDA/IRF4/SLC24A4.
Key terms — so this notebook stands on its own (you shouldn’t need the other notebooks to read this one).
- Effector-uncertain locus — a pigmentation-associated genome location where which gene actually drives the trait is genuinely unknown: the nearest gene is NOT a textbook pigment gene. These 18 loci are the “rescue” targets this diagnostic tests.
- Canonical pigmentation gene / effector (and “canonical neighbor”) — a gene with a textbook melanin-making role (TYR, OCA2, MC1R, SLC45A2, MITF, TYRP1, SLC24A4…). A locus that connects “only via a canonical neighbor” is explained by the known gene already sitting there, not by anything the screen discovered.
- Melanogenesis network / “substrate” / core — the pathway of genes that make melanin, rebuilt in earlier notebooks as a signed (activation/repression), directed graph; here it is the 803-node / 7,819-edge “substrate” graph (from NB7), whose edges are tiered by confidence (T0 Raghunath curation → GRN regulon → OmniPath → STRING). A locus is “rescued” if its candidate gene reaches this core.
- Rescue screen & “non-circular route” — the project aim: take a locus whose effector is unknown and try to link it into the melanogenesis network, grading by how many independent evidence lines converge. “Non-circular” means a genuine new link (direct substrate edge, GRN regulon, Reactome subpathway, melanocyte eQTL, ≤2-hop path, or verified biology) rather than just re-using the canonical gene already at the locus.
- GRN / MITF-SOX10-PAX3 regulon — gene-regulatory network: transcription-factor→target edges, signed activation/repression. MITF, SOX10, and PAX3 are the master pigment transcription factors; their “regulon” is the set of genes they regulate (58 signed edges, built in NB6).
- Melanocyte vs bulk-skin eQTL, eGene, coloc (CLPP) — an eQTL is a variant linked to a gene’s expression level; an eGene is a gene that has one. Colocalization (coloc) tests whether a GWAS signal and an eQTL share the same causal variant, scored by CLPP (probability; 1.0 = certain shared signal). Using melanocyte-specific data (Zhang 2018) instead of bulk skin gives the right cell type, so it can retract a wrong causal-gene call or confirm a right one.
- L2G (locus-to-gene) — Open Targets’ score for which gene a GWAS locus most likely acts through; the “provisional bulk L2G resolution” is the pre-melanocyte guess this notebook re-checks.
- Convergence grade (A/C) — how much independent evidence a link carries: grade A = several independent lines agree (e.g., LRMDA); grade C = essentially one line. Evidence is only ever ADDED, never used to drop a locus.
- STRING / OmniPath — STRING is a database of functional protein–protein associations (correlation, not mechanism or direction); OmniPath adds curated signed/directed interactions. These are the lowest-confidence graph tiers, so a link resting on STRING alone does not license a mechanistic claim.
- Bajpai 2023 CRISPR / casTLE — a genome-wide CRISPR-knockout screen for melanin content (casTLE = its effect-size statistic). A “hit” corroborates a candidate gene; a non-hit is NOT evidence against it (the screen used one cell line).
import pandas as pd, numpy as np, json
from pathlib import Path
pd.set_option('display.max_colwidth', 60)
REPO = str(Path.cwd().parent if Path.cwd().name == 'notebooks' else Path.cwd())
diag = pd.read_csv(f'{REPO}/data/processed/nb8_diagnostic_18.csv')
print('18 loci:', diag.shape)
diag['bucket'].value_counts()data/processed/provisional_effector_uncertain_18.csv (Ang2023 Kalinago albinism sib-set + Morgan2018 UKB).nb7_substrate_nodes.csv (803) / nb7_substrate_edges.csv (7819; tiers T0_Raghunath, T1_GRN_regulon, T2b_OmniPath, T3_STRING).nb6_grn_edges.csv (58 signed MITF/SOX10/PAX3 regulon edges; 45 targets).reactome_mitf_subpathway_structure.json — pigmentation subpathway R-HSA-9824585 has 29 genes in our 121-set (gated tag), NOT “N of 121”..xlsx — T-S6 (379 curated pigment genes w/ melanocyte eGene q), T-S9 (trans-eQTL), T-S10 (IRF4-mediated trans: NEO1/PLA1A/TMEM140), T-S15 (melanocyte colocalization of pigment-trait GWAS).# core connection screen was run per-locus across all 6 evidence lines.
# See nb8_diagnostic_18.csv for the per-locus route/edges/tiers/grade/bucket/experiment.
for _,r in diag.iterrows():
print(f"[{r['bucket']:8}] {r['rsid'][:22]:22} nearest={str(r['nearest_gene_label'])[:26]:26} -> route={r['connection_route'][:38]}")The melanocyte layer changes the call at several loci relative to the provisional bulk-skin L2G resolution: - RETRACTED (bulk gave a gene; melanocyte data does not support it — gene is not a melanocyte eGene): TECRL (locus4), RAB11FIP2 (locus9), CCND1 (locus16, bulk-only). - CONFIRMED / STRENGTHENED: LRMDA=C10orf11 melanocyte eGene q=3.9e-11 (locus1); IRF4 melanocyte coloc CLPP=1.0 (locus6); SLC24A4 melanocyte coloc CLPP=0.46 (locus11). - IRF4 trans-eQTL (Zhang T-S10) drives NEO1/PLA1A/TMEM140 — but those are downstream of canonical IRF4, not a rescue of the uncertain locus.
from IPython.display import Image
Image(filename='figures/nb8_evidence_matrix.png') # 18 loci x 6 evidence lines; teal=non-canonical, grey=canonical/generic
Regulation of MITF-M-dependent genes involved in pigmentation tag; the other 92 are MITF-program-only and are reported separately, never merged.