Gene-network datasets inventory — `tinalasisi/pigmentation-gene-network`

Gene-network datasets inventory — tinalasisi/pigmentation-gene-network

Purpose. SEPIA item B1 prep: catalog every dataset the public repo tinalasisi/pigmentation-gene-network provides, so we can plan folding its richer, gene-level pigmentation data into SEPIA (the "bigger goal" beyond Lily's 222-SNP panel).

Provenance of this inventory. Cloned https://github.com/tinalasisi/pigmentation-gene-network at commit 7618746 (2026-07-13). Every count, column, DOI, and build below was read directly from the repo's committed DATA_SOURCES.md and by inspecting each file's header + rows. Where a figure below differs from DATA_SOURCES.md, the file-actual value is given and the difference noted. Nothing here is inferred beyond what was read; missing items are called out as missing.

Build reality (read this first — it is not "all hg38")

DATA_SOURCES.md and SEPIA plan.md §5 both describe the repo as "hg38." That is true for the two pieces SEPIA most cares about, but not uniform across the repo:

• hg38 (GRCh38): the GWAS Catalog pull (pos_hg38 column, 1,065/1,072 rows populated) and the assembled gene network's gene coordinates (MyGene genomic_pos, default GRCh38). These are the datasets SEPIA calls the "four hg38 gene-network datasets."
• hg19 (GRCh37): the per-paper SNP extractions carry coord_build = GRCh37/hg19 and pos_hg19 columns — Martin 2017, Kim 2024, the discordance/case-study loci, and the merged nb4 base's curated-paper rows. In nb4_unified_association_base.csv the two builds sit side by side: 1,072 rows tagged GRCh38 (pos_hg38 … GWAS Catalog) and 105 curated-paper rows mostly GRCh37/hg19 (b37).
• build-agnostic: HIrisPlex-S, Bajpai, Baxter, Raghunath, and the D'Arcy tables are keyed by rsID or gene symbol only — no genomic coordinates at all (except the assembled network's gene chr), so "build" does not apply until coordinates are attached.

Implication for SEPIA's B2 bridge: rsID is the only field common to all SNP-level sources, which is exactly why the B2 plan keys the hg19↔︎hg38 map on rsID. Coordinate columns in these files cannot be trusted to be hg38 without checking the per-file coord_build tag.

Source index (what the repo provides)

The five SEPIA §5 "four datasets + prediction panel" map to rows above: GWAS Catalog (1/1b), CRISPR = Bajpai (2), cross-species = Baxter (3), melanogenesis network = Raghunath (5), PPI expansion = D'Arcy (6/6b/6c), prediction panel = HIrisPlex-S (4).

Per-dataset detail

1 · GWAS Catalog — pigmentation associations (SNP-level, hg38)

• What: NHGRI-EBI GWAS Catalog lead SNPs for pigmentation (skin/eye/hair + photo-response), pulled over 10 frozen EFO/OBA/MONDO trait roots with child traits.
• Provenance: live pull, download endpoint https://www.ebi.ac.uk/gwas/api/search/downloads, queried_utc = 2026-07-12T14:46:57Z; script scripts/gwas_catalog.py v1.0. Catalog citation Sollis 2023; EMBL-EBI terms (reuse with attribution) + per-row study PubMed IDs. Independently cross-checked via the human-genetics MCP connector.
• File / format: pigmentation_gwas_catalog.csv, 27 columns including rsid, chr, pos_hg38, risk_allele, direction_raw, risk_freq, or_beta, pvalue, mapped_gene, reported_gene, pubmed, study_accession, ancestry, initial/replication ancestry+N, queried_utc. Build GRCh38/hg38 (pos_hg38; 1,065 of 1,072 rows populated — 7 rsIDs unmapped).
• Counts: 1,072 rows = 1,072 unique rsIDs. Payoff loci per DATA_SOURCES.md: MC1R 16 assoc · OCA2 67 · HERC2 65. p-values kept as strings (underflow-safe); no p-filter at pull.
• 1b (granular sibling): gwas_pigmentation_associations.csv — 723 rows, columns efo_id, trait, gene, snp_id, pvalue; 472 unique rsIDs. This is the per-gene / per-association table (supports the ≥2-association replication filter → 83 replicated genes); the deduplicated 1,072 table cannot give per-gene counts.

2 · Bajpai 2023 — genome-wide CRISPR screen (gene-level)

• What: 169 genes whose CRISPR perturbation changed melanin content in a genome-wide screen (Table S1, "Low SSC FACS enriched genes"), called at q < 0.10 (reproduces the paper).
• Provenance: DOI 10.1126/science.ade6289 · PMC10901463 · Science 381(6658):eade6289. Table S1 is a supplementary file of the CC BY 4.0 PMC deposit (redistributable w/ attribution).
• File / format: bajpai2023_crispr_hits.csv, 13 columns — GeneID(Ensembl), Symbol, GeneInfo, Localization, Process, Function, Combined_casTLE_Effect, Combined_casTLE_Score, Minimum/Maximum_Effect_Estimate, p_value, q_value, direction_note.
• Counts: 169 gene rows, no rsIDs (verified 0 rsID matches). All effects positive → uniform direction perturbation reduces pigmentation. TYR/SLC45A2/OCA2 are hits; MC1R/HERC2 are not.

3 · Baxter 2018/19 — curated cross-species gene list (gene-level)

• What: curated pigmentation genes with mouse/zebrafish orthologs (Supporting Table S7, "650 Pigmentation Genes").
• Provenance: DOI 10.1111/pcmr.12743 · PMID 30339321 · PMC10413850 · Pigment Cell Melanoma Res. Raw Table S7 is NOT redistributed (Wiley subscription VOR, no open-license copy; git-ignored); only the derived CSV is committed. Cite Baxter + OMIM/MGI/ZFIN/GO.
• File / format: baxter2018_650_pigmentation_genes.csv, 12 columns — Gene stable ID, Human gene symbol, Mouse gene symbol, Zebrafish gene symbol, Orthologs across species, Pigment phenotype location, GO, OMIM, MGI, ZFIN, PubMed, Species with phenotype.
• Counts: 659 rows; 635 carry a human gene symbol (the project's "Baxter 635"). Membership list only — no direction, no effect size, no rsIDs. OCA2, MC1R present; HERC2 absent.

4 · HIrisPlex-S — Chaitanya 2018 forensic prediction markers (SNP-level) ★

• What: the forensic skin/hair/eye-color prediction marker set.
• Provenance: DOI 10.1016/j.fsigen.2018.04.004 · FSI:Genetics 35:123–135; skin-model coefficients from Walsh 2017 (Hum Genet 136:847); web tool hirisplex.erasmusmc.nl. rsIDs are factual; model free to use via the Erasmus MC tool. Elsevier subscription for the PDF text. Extraction notebook notebooks/01c_extract_hirisplex_markers.ipynb; raw transcription data/raw/hirisplexs2018/hirisplexs2018_markers_raw_transcribed.csv.
• File / format: hirisplexs2018_markers.csv, 3 columns — gene, rsid, in_novel_17plex_skin.
• Counts: 36 markers = 36 unique rsIDs → 16 genes. DATA_SOURCES.md documents this as 36/41 (full 41 needs Table 1 transcription; one OCR fix rs228479→rs2228479, one dup dropped).
• The 16 genes: ANKRD11, ASIP, BNC2, DEF8, HERC2, IRF4, KITLG, MC1R, OCA2, PIGU, RALY, SLC24A4, SLC24A5, SLC45A2, TYR, TYRP1.
• MC1R red-hair set — complete (8 markers present): rs1805007, rs1805008, rs1805006, rs11547464, rs885479, rs2228479, rs1110400, rs3212355.
• Blue-eye markers — present: HERC2 rs12913832 ✓ and OCA2 rs1800407 ✓ (also OCA2 rs1800414, rs1470608, rs1545397, rs12441727; HERC2 rs2238289/rs6497292/rs1129038/rs1667394/rs1667395).
• Companion hirisplexs2018_population_provenance.json documents the model training populations (eye ~97% European, hair ~85% European, skin most diverse) — relevant to SEPIA's European-training caveat for archaic prediction. Numeric prediction weights are NOT in this repo (must be pulled from Walsh 2017 / the web tool) — matches SEPIA plan's "optional fetch" note.

5 · Raghunath 2015 — directed melanogenesis backbone (gene/network)

• What: the mechanistic directed/signed melanogenesis network; the backbone everything else annotates.
• Provenance: BMC Res Notes 2015, Additional Files 1–5 (CC BY 4.0; raw MOESM1–5 committed in data/raw/raghunath2015/). Prior in-lab SEPIA work.
• Files / format: raghunath_nodes_typed.csv (cols node, base, node_type, type_source, compartment, state, dual_compartment) + raghunath_edges_typed_signed.csv (cols source, target, verb, interaction_norm, sign, reference).
• Counts: 265 nodes / 429 signed directed edges. Nodes are a mix of genes/proteins and metabolites (e.g. 4HNE_kerat, ACTH_kerat) with compartment tags — not a clean gene list. No rsIDs. OCA2 present; HERC2 absent; MC1R present in both keratinocyte and melanocyte compartments.

6 · D'Arcy 2023 — OMIM disease genes + STRING PPI + mass-spec (gene/network)

• Provenance: DOI 10.3390/bioengineering10010013 · PMC9854651 · PMID 36671585 · Bioengineering 10(1):13. CC BY 4.0, six supplementary tables committed at data/raw/darcy2023/Table S1–S6 …FINAL.xlsx; derived CSVs in data/processed/darcy2023_S*.csv (each with a .meta.json).
• Table S1 — disease genes: darcy2023_S1_disease_genes.csv, cols gene, disease_name, inheritance, phenotype_mim, phenotype_class, pigmentation_phenotype, source, citation, citation_source. 278 rows / 243 unique genes (DATA_SOURCES.md calls it the "243-gene OMIM disease-gene table" — the extra rows are multiple disease entries per gene; S2 sys/GO annotation confirms 243 unique genes).
• Tables S4/S5 — STRING PPI: darcy2023_S4_string_edges.csv (4,668 edges, cols node1, node2, combined_score, citation, citation_source) + darcy2023_S5_string_nodes.csv (452 nodes, cols incl. gene_string, process_1..3, main_location, disease_gene_flag/class, in_a375, in_fm55). STRING combined_score, undirected/unsigned association — not mechanistic. (DATA_SOURCES.md states "451-node/4668-edge"; file has 452 node rows.)
• Table S6 — mass-spec: darcy2023_S6_massspec_expression.csv, 4,232 genes, A375/FM55 melanoma LFQ expression. Table S2 — sys/GO: 243 genes annotated.
• No rsIDs anywhere in the D'Arcy tables.

7 · Assembled gene network — Notebook 2 output (gene/network, hg38 gene coords)

• data/processed/gene_network_nodes.csv (168 nodes, cols gene, entrez, ensembl, chr, node_class, citation, citation_source) + gene_network_edges.csv (309 edges, cols source, target, sign, edge_class, via, citation, citation_source). Gene identity via MyGene/UniProt (GRCh38 chr), release-gated so every node+edge carries a resolvable citation. No rsIDs. This is the repo's own integrated backbone (Raghunath + database enrichment); downstream NB5 layers Bajpai/Baxter/D'Arcy onto it.

Additional SNP-level sources (present, hg19 — beyond the SEPIA §5 five)

These are extra rsID-bearing tables the repo also ships. They are hg19/GRCh37 and are relevant to SEPIA only if the panel is extended beyond the §5 set:

• EXTRACT_Martin2017_loci.csv — KhoeSan (San/Nama) skin-pigmentation GWAS, 51 rows / 49 unique lead rsIDs, pos_hg19, San/W-African/N-European allele freqs. DOI 10.1016/j.cell.2017.11.015.
• EXTRACT_Kim2024_loci_v2.csv — East-Asian skin-color GWAS, 26 rows / 26 rsIDs, coord_build = GRCh37/hg19. DOI 10.1038/s41467-024-49031-4.
• discordance_loci.csv — 105 rows / 75 unique rsIDs (eye-color discordance case studies), hg19.
• data/case_records/EXTRACT_*_records.csv (13 papers) + data/processed/EXTRACT_*_loci_v2.csv (15 papers) — per-individual validation case-study genotype records (Norton TYRP1, Morgan MC1R, Yang OCA2, etc.). SNP/genotype-level, hg19, small.
• nb4_unified_association_base.csv — the merged SNP base: 1,177 rows / 1,103 unique rsIDs = 1,072 GWAS-Catalog-dedup (hg38) + 105 curated-paper rows (mostly hg19), with rsid, rsid_primary, pos_hg38, coord_build, gene_label, pvalue, population, and cross-membership flags. This is the single most convenient union table if SEPIA wants "all pigmentation rsIDs the repo knows about."

Integration into SEPIA

SEPIA's plan (plan.md B1/B3/B6) already splits sources into SNP-level (feed the directional polygenic score B3 + panel PCA) vs gene-level (gene-region PCA B4-ii, layered coverage B1, functional interpretation B7). This inventory confirms and quantifies that split.

A. rsID SNP-lists → fold into the hg19↔︎hg38 panel map (B2), extend the panel (B1/B3)

These have rsIDs, so B2's rsID ↔︎ hg19 ↔︎ hg38 map can resolve them and we can pull archaic/SGDP genotypes at each locus and report on hg38:

Overlap method / caveat: SEPIA's 222 panel positions (data/snps.txt, hg38) map cleanly to 222 rsIDs via data/skin_pigmentation.tsv (the panel's GWAS-Catalog source table, 226 unique rsIDs). Overlap counts above are rsID-set intersections against those 222 (identical whether measured against the 222-mapped set or the 226-rsID source table, ±1). So:

• The GWAS Catalog pull is a strict superset of the current panel (212/222 present) and roughly quadruples the pigmentation SNP count (→ ~1,072). The 10 panel rsIDs not found are likely non-pigmentation/near-miss trait rows dropped by the repo's trait-root filter — worth a spot check before swapping the panel source.
• HIrisPlex-S contributes 21 forensically-validated markers not in the panel, including the well-characterized MC1R red-hair and HERC2/OCA2 blue-eye variants SEPIA's Q9 concordance check (plan.md B7) explicitly wants. This is the highest-value small addition.
• Martin/Kim add non-European ancestry axes but are hg19 — they still route through the same rsID bridge, so no extra machinery; include only if the phenotype scope broadens.

Build handling: fold these in by rsID, not by coordinate. Only the GWAS Catalog rows are natively hg38; HIrisPlex-S is rsID-only and Martin/Kim are hg19. B2's dual-coordinate dbSNP re-pull (keyed by rsID) is exactly the right mechanism, and nb4_unified_association_base.csv is a ready-made union to seed the map (it already stores rsid+pos_hg38+coord_build per row).

B. Gene-level → B6 (genes-vs-SNPs) / B4-ii (gene-region PCA) / B7 (functional)

No per-SNP genotype; used as gene-region sets (take each gene's Ensembl start/end, use all variants inside) or for network/functional interpretation — never in the +1/−1 directional score:

• Bajpai 169 CRISPR hits — gene set + Combined_casTLE_Effect weight, uniform reduces-pigmentation sign.
• Baxter 635 curated genes — membership only.
• Raghunath 265-node / 429-edge signed directed backbone — mechanism/direction context (nodes include metabolites; filter to gene nodes before coordinate lookup).
• D'Arcy 243 disease genes + 452-node/4,668-edge STRING PPI (association, not mechanistic) + 4,232-gene mass-spec — disease-gene overlay + PPI-shell expansion for B7.
• Assembled gene network (168 nodes, GRCh38 chr) — the repo's integrated backbone, already citation-gated.

These need gene coordinates from Ensembl/BioMart on the matched build (SEPIA plan B4-ii/B6 already specify this); the repo supplies symbols/Ensembl IDs, not coordinates (except the 168-node assembled network's chr), so a coordinate join is required regardless.

C. What is NOT in the repo (do not assume)

• HIrisPlex-S numeric prediction weights — absent; pull from Walsh 2017 / Erasmus MC tool if quantitative color probabilities are needed (SEPIA already flags this as optional for archaics).
• Per-SNP genotypes / allele-frequency panels for archaics — not here; that is SEPIA's own data.
• Kim 2024 / Zhang 2018 full summary stats — only reference/derived extracts; full stats are publisher/dbGaP-gated.
• Coordinates for the gene-level sources — symbols only.