GWAS Mechanisms: Gene-Program-Trait Causal Modeling

Overview

Use DisMech as a validation/interpretation layer for computational pipelines that discover causal gene-to-trait relationships via GWAS, Perturb-seq, and causal modeling. The pilot application is the Ota et al. (Nature 2025) framework that builds three-layer causal graphs:

Gene (regulator) --[β]--> Program --[effect]--> Trait/Phenotype
                          ↑
Gene (member) ----[γ]----/

This project covers: (1) validating pipeline outputs against curated pathophysiology, (2) curating blood/hematological disorders to improve coverage, (3) building tooling to map between pipeline entities and dismech ontology terms, and (4) extending to other cell types and trait domains as new Perturb-seq datasets emerge.

Background: Ota et al. (Nature 2025)

Paper: "Causal modelling of gene effects from regulators to programs to traits" - Ota M, Spence JP, Zeng T, Dann E, Milind N, Marson A, Pritchard JK - Nature 650:399-408 (Feb 2026; online Dec 10, 2025) - DOI: 10.1038/s41586-025-09866-3 - PMID: 41372418 | PMCID: PMC12893915 - bioRxiv: 2025.01.22.634424v1

Key methodology: - 60 gene programs from consensus NMF on K562 Perturb-seq (Replogle et al. 2022; 9,866 genes) - Gene-trait effect sizes from UK Biobank LoF burden tests (454K participants) via GeneBayes - S-LDSC to identify traits where K562 chromatin explains heritability - Multiple regression modeling gene effects on programs and program effects on traits - 73% accuracy predicting effect directions for top MCH GWAS hits

Key biological findings: - SUPT5H regulates hemoglobin production, cell cycle, and autophagy simultaneously - Programs capture biologically meaningful co-regulation modules - Regulators of a program vs. member genes can have distinct trait relationships

Related Work

Assets

External Code Repositories

• GeneBayes - empirical Bayes framework (Zeng et al.)
• specificity_length_luck - companion paper code
• GWT_perturbseq_analysis_2025 - T cell follow-up
• No dedicated repo for the main Ota et al. analysis code (as of Feb 2026)

Existing Blood/Hematological Disorders in KB

Key Programs Relevant to DisMech Disorders

Traits (54 total from Table S2)

RBC Traits (12) - primary focus

Reticulocyte count, Mean reticulocyte volume, Mean sphered cell volume, Immature reticulocyte fraction (IRF), High light scatter reticulocyte count, RBC count, MCV, MCH, MCHC, RDW, Haematocrit, Haemoglobin concentration

Other Blood Traits (10)

WBC count, Platelet count, Platelet crit, MPV, Platelet distribution width, Lymphocyte count, Monocyte count, Neutrophil count, Eosinophil count, Basophil percentage

Serum Biomarkers (28)

Albumin, ALP, ALT, ApoA, ApoB, AST, Direct bilirubin, Urea, Calcium, Cholesterol, Creatinine, CRP, Cystatin C, GGT, Glucose, HbA1c, HDL, IGF-1, LDL, Oestradiol, Phosphate, SHBG, Total bilirubin, Testosterone, Total protein, Triglycerides, Urate, Vitamin D

Anthropometric (4)

Sitting height, Birth weight, BMI, Heel BMD

Cross-reference: PGS×context amplification (pgs_context_amplification)

A separate line of curation in this repo encodes the PGS-by-context (PGS×C) amplification hypothesis from Nagpal & Gibson (Nat Genet 2026, PMID:42443528) as disease-level mechanistic_hypotheses blocks that all share hypothesis_group_id: pgs_context_amplification. Conformers so far: Coronary_Artery_Disease, Type_2_Diabetes_Mellitus, Chronic_Kidney_Disease, Obesity, Asthma, Crohn_Disease, Ulcerative_Colitis. Each proposes that common-variant polygenic liability and an adverse exposure are amplified because they converge on a shared pathophysiology node (e.g. Endothelial Dysfunction for CAD, Airway Inflammation for asthma, Mucosal Inflammation for UC), and each carries a paired reverse-causation discussions KNOWLEDGE_GAP.

Why it connects to this project. The two efforts are orthogonal — Ota et al. dissects the gene→trait path holding environment fixed; PGS×C varies the environment holding the path implicit — but they share a hinge, the dismech pathophysiology node:

• Ota et al. supplies the gene → program → trait half experimentally (Perturb-seq β, rare/LoF-burden γ); an Ota program maps onto a dismech pathophysiology node.
• The PGS×C work supplies the environment → node half observationally and names the convergence node where common-variant genetic liability and the exposure collide.

An Ota program is therefore a candidate molecular substrate for a pgs_context_amplification convergence node: if genetic risk acts through program P and an exposure perturbs program P, that shared program is the mechanism for the amplification the PGS×C hypothesis currently asserts only statistically — and Perturb-seq would give it the interventional grounding the observational PGS×C analysis lacks (the reason each PGS×C hypothesis ships with a reverse-causation gap). Concrete candidate pairings from the program table above: P46 (cholesterol biosynthesis) ↔ the Coronary_Artery_Disease convergence node; P16 (autophagosome) and P35/P45 (TNF/NF-κB) ↔ the Crohn_Disease / Ulcerative_Colitis inflammation convergence nodes. Asthma and Ulcerative_Colitis already appear in this project's disease list, so those two entries now carry both layers at once.

Naming caveat: the amplification concept the PGS×C hypotheses cite (PMID:37228747 — Zhu et al., gene-by-sex amplification, Harpak lab) is a different paper from the "Zhu, Dann, …, Ota" CD4+ T cell Perturb-seq follow-up in the Related Work table above, despite the shared lead-author surname.

TASKS

Phase 1: Blood Disorder Curation Gap-Fill

Priority: curate missing blood disorders that are highly relevant to the 12 RBC traits.

• [ ] Alpha Thalassemia
• [ ] Beta Thalassemia
• [ ] Iron Deficiency Anemia
• [ ] Aplastic Anemia
• [ ] Myelodysplastic Syndromes
• [ ] Diamond-Blackfan Anemia
• [ ] Congenital Dyserythropoietic Anemia
• [ ] Hereditary Spherocytosis

Phase 2: Trait-to-HPO Mapping

Map all 54 traits from Table S2 to HPO terms (increased/decreased variants).

• [ ] Create trait_to_hpo.yaml mapping file for RBC traits (12)
• [ ] Create mappings for other blood traits (10)
• [ ] Create mappings for serum biomarkers (28)
• [ ] Create mappings for anthropometric traits (4)

Example mapping needed:

MCH:
  trait_name: "Mean corpuscular haemoglobin"
  lof_trait_id: 30050
  hpo_decreased: HP:0025066  # Decreased mean corpuscular hemoglobin
  hpo_increased: HP:0025065  # Increased mean corpuscular hemoglobin
  related_disorders:
    - Sickle_Cell_Disease
    - Beta_Thalassemia

Phase 3: Program-to-GO Mapping

Map all 60 programs to proper GO terms (many currently use Hallmark gene set names).

• [ ] Resolve HALLMARK_* annotations to GO terms
• [ ] Validate GO terms via OAK
• [ ] Create program_go_mapping.yaml

Phase 4: Proof-of-Concept Validation (MCH/RDW/IRF)

Manual validation of top gene-program-trait relationships for the three blood traits.

• [ ] Extract top 20 genes per trait from paper/supplements
• [ ] For each gene-program-trait triple, check if dismech documents the relationship
• [ ] Classify as CONFIRMED / PARTIAL / NOVEL / CONTRADICTED
• [ ] Write validation report

Phase 5: Automated Validation Pipeline

Build tooling to systematically match pipeline outputs against dismech.

• [ ] Export dismech blood disorders to queryable format (gene, GO, HP indexed)
• [ ] Implement gene-program-trait matching logic
• [ ] Generate validation report with coverage statistics
• [ ] Identify curation gaps from NOVEL findings

Phase 6: Extend to T Cell / Immune Traits

Leverage the Zhu/Dann et al. CD4+ T cell Perturb-seq follow-up.

• [ ] Obtain T cell program definitions when published
• [ ] Map immune traits to HPO terms
• [ ] Validate against dismech autoimmune disease entries (20+ curated)
• [ ] Compare K562 vs T cell program overlap
• [ ] Cross-check pgs_context_amplification convergence nodes (Asthma, Ulcerative_Colitis, Crohn_Disease) against T cell / immune programs as candidate molecular substrates for the amplification (see "Cross-reference: PGS×context amplification" above)

Phase 7: Schema Enhancements

Address dismech schema gaps identified in the integration plan.

• [ ] Add systematic modifier (INCREASED/DECREASED) to phenotype entries
• [ ] Add genes to pathophysiology biological_process entries where missing
• [ ] Consider gene slot on BiologicalProcessDescriptor for explicit gene-process links

NOTES

Zhu/Dann T Cell Follow-up: Traits and Diseases

The T cell paper takes a different approach from the K562 paper. Rather than directly modeling continuous GWAS traits, it tests:

1. Cytokine readouts (30 genes): IFN-gamma, TNF, IL-4, IL-5, IL-13, IL-10, IL-2, IL-16, IL-21, CCL3, CCL4, CCL5, CXCL8, TGF-beta, etc.
2. Th1/Th2 polarization signatures from published bulk RNA-seq
3. CD4+ T cell aging signatures from OneK1K cohort (782 donors)
4. Lymphocyte counts from UK Biobank (Backman et al. 2021 exome data)
5. GWAS gene enrichment for 14 autoimmune diseases via Open Targets

Autoimmune diseases tested: | Disease | Open Targets ID | In dismech KB? | |---------|----------------|----------------| | Rheumatoid arthritis | EFO_0000685 | Yes | | SLE | MONDO_0007915 | Yes | | IBD | EFO_0003767 | Yes (Crohn's + UC) | | Multiple sclerosis | MONDO_0005301 | Yes | | Type 1 diabetes | MONDO_0005147 | Yes | | Psoriasis | EFO_0000676 | Yes | | Ankylosing spondylitis | EFO_0003898 | Yes | | Asthma | MONDO_0004979 | Yes | | Hashimoto's thyroiditis | EFO_0003779 | Yes | | Crohn's disease | EFO_0000384 | Yes | | Ulcerative colitis | EFO_0000729 | Yes | | Celiac disease | EFO_0001060 | Yes | | Atopic eczema | EFO_0000274 | Yes (Atopic Dermatitis) | | Autoimmune disease (general) | EFO_0005140 | — |

Key findings: - 111 regulator clusters from 3,341 perturbations of 1,860 regulators - Cluster 110 (12 genes incl. MAP3K8, CHD7, ITK): massive enrichment for asthma (OR=20.4), atopic eczema (OR=24.3), psoriasis (OR=17.6) - Cluster 79 (Th17; IRF4, BATF, STAT3, IPMK): Crohn's (OR=58.2), eczema (OR=38.1) - Cluster 59 (IL10, BACH2, REL, CTLA4, CD28): broad enrichment across IBD, psoriasis, MS

Key methodological difference: Uses GWAS gene enrichment per cluster (via Open Targets) rather than direct trait-effect modeling. Also uses pert2state models for aging/polarization.

GitHub repo: https://github.com/emdann/GWT_perturbseq_analysis_2025 - Supplementary tables in metadata/suppl_tables/ - Full clustering results: clustering_results_and_annotations.csv - Autoimmune enrichment: cluster_autoimmune_enrichment_results.suppl_table.csv

Blood Disorder KB Readiness Assessment

Best candidates for K562/blood validation: CML, PV, ET, PMF (myeloproliferative neoplasms) Best candidates for T cell/immune validation: All 14 autoimmune diseases already in KB

2026-02-14 - T cell/autoimmune validation completed

Result: 2.7% gene confirmation rate (11/408 genes across 12 diseases, 146 significant pairs).

This reflects a gene coverage gap, not contradictions. dismech has 3-11 genes per disease (the classic GWAS hits) while the pipeline tests hundreds from Open Targets. The confirmed genes (PTPN22, CTLA4, IL2RA, IL4, IL23R, etc.) are exactly the well-documented ones.

Key outputs: - Validation report: docs/gwas-tcell-validation-report.md - Validation script: scripts/validate_tcell_clusters.py - Data: docs/assets/zhu_dann_2025/

Top curation targets identified: EGR2, BACH2, ETS1, IRF4, TNFAIP3, IKZF1, CD28, STAT3, GATA3, SMAD3, IL10 -- all well-known autoimmune genes missing from dismech.

Cluster 79 (Th17: IRF4, BATF, STAT3, JUNB) is the standout -- OR=58.2 for Crohn's, 38.1 for eczema, 26.9 for psoriasis. Th17 pathophysiology should be strengthened across entries.

Missing KB entry: Atopic Dermatitis/Eczema (16 significant pairs, strong pipeline signal).

Next steps: Add pleiotropic autoimmune genes to relevant entries; curate Atopic Dermatitis; strengthen Th17 pathophysiology in Crohn's, Psoriasis, Ankylosing Spondylitis.

2026-02-14 - Mechanism-level gap filling

Addressed GO term and cell type gaps across autoimmune entries to enable mechanism-level matching against Zhu/Dann T cell clusters.

Asthma.yaml (most significant changes): - Added biological_processes GO terms to ALL 4 existing pathophysiology entries: - Airway Inflammation: GO:0006954 inflammatory response, GO:0002437 inflammatory response to antigenic stimulus - Bronchoconstriction: GO:0006939 smooth muscle contraction - Mucus Overproduction: GO:0070254 mucus secretion - Airway Remodeling: GO:0048771 tissue remodeling - Added NEW pathophysiology entry: "Type 2 Immune Response / Th2 Signaling" - 5 GO biological_processes: GO:0042092, GO:0045064, GO:0035771, GO:0035708, GO:0032633 - 6 genes: IL4, IL13, IL4R, STAT6, IL5, GATA3 (with HGNC IDs) - 4 cell types: Th2 cell, ILC2, mast cell, eosinophil - Downstream links to Airway Inflammation and Mucus Overproduction - This directly enables matching against Cluster 85 (STAT6, IL4R, RBPJ)

Psoriasis.yaml: - Added GO:0072538 (T-helper 17 type immune response) to IL-23/IL-17 Axis entry - Added cell types (Th17 cell, dendritic cell) that were missing

Ankylosing_Spondylitis.yaml: - Added GO:0072538 to IL-23/IL-17 Axis Activation entry - Changed cell type from generic CD4+ T helper to specific T-helper 17 cell - Added dendritic cell

Rheumatoid_Arthritis.yaml: - Added biological_processes to Autoimmune Response (had Th17 cell type but zero GO terms) - GO:0072538 (T-helper 17 type immune response) + GO:0006954 (inflammatory response)

Multiple_Sclerosis.yaml: - Added NEW pathophysiology entry: "Th1/Th17-Mediated Neuroinflammation" - GO:0072538 (Th17 immune response), GO:0006954 (inflammatory response) - Cell types: T-helper 17 cell, T-helper 1 cell - Downstream link to Inflammatory Lesions - Evidence from PMID:32801039 (Moser et al. 2020) and PMID:21338381 (Jadidi-Niaragh 2011) - Added biological_processes + cell_types to Inflammatory Lesions entry

Systemic_Lupus_Erythematosus.yaml: - Added biological_processes to Formation of Immune Complexes (GO:0002377, GO:0006958, GO:0019724)

Hashimotos_Thyroiditis.yaml: - Added biological_processes to Lymphocytic Infiltration (GO:0006954, GO:0002250)

Ulcerative_Colitis.yaml: - Added biological_processes to Dysregulated Immune Response (GO:0042092, GO:0006954)

NEW: Atopic_Dermatitis.yaml (complete new entry): - 4 pathophysiology mechanisms with GO terms, cell types, evidence - 5 phenotypes with HP terms, 5 genetic entries, 3 environmental factors, 5 treatments - Evidence from PMID:16550169, PMID:30819278, PMID:21388665 - Enables matching against Cluster 79 (Th17, OR=38.1) and Cluster 85 (Th2)

Validation improvement: 13 diseases evaluated (up from 12), 162 cluster-disease pairs (up from 146)

Remaining gaps: - [x] Add pleiotropic autoimmune genes (EGR2, BACH2, ETS1, IRF4, STAT3, etc.) across entries → DONE (see below) - [ ] Epigenetic regulation mechanism (clusters 38, 59) - not documented in any entry - [ ] T cell metabolic reprogramming / mTOR (cluster 100) - undocumented

2026-02-14 - Pleiotropic gene additions (batch 1: 6 genes)

Added BACH2, TNFAIP3, STAT3, IL10, CD28, EGR2 across 12 disease files. Result: 17 confirmed genes, 4.0% rate.

2026-02-14 - Pleiotropic gene additions (batch 2: 10 genes)

Added 10 more pleiotropic GWAS genes across all 12 autoimmune disease files: - ETS1 (10 diseases), IRF4 (9), IRF8 (7), SATB1 (8), IKZF1 (8) - SMAD3 (8), REL (6), PRDM1 (6), PTPN22 extension (6 new), IL21R (4)

Result: 26 confirmed genes, 6.1% rate (up from 4.0%)

Per-disease rates: Celiac 29.6%, Psoriasis 25.0%, MS 19.1%, SLE 19.4%, T1D 18.9%, AS 17.6%, Crohn's 15.3%, AD 15.2%, UC 15.5%, Hashimoto's 12.8%, RA 12.0%, Asthma 10.6%

Top remaining novel genes (diminishing returns): IL2RA extension, FOSL2, IL7R, ADO, FADS2, BCL2L11, PRKCB - these are less established classical autoimmune GWAS hits.

Remaining mechanistic gaps: - [x] Epigenetic regulation (clusters 38/59) → Added to RA and SLE - [x] T cell metabolic reprogramming (cluster 100) → Added to SLE (combined entry) - [ ] IL2RA extension to remaining diseases (already in T1D, needs 8+ more) - [ ] Additional epigenetic entries could be added to MS, Crohn's, psoriasis

2026-02-14 - Epigenetic/metabolic mechanism entries

Added two new pathophysiology entries to capture cross-cutting regulatory mechanisms from Zhu/Dann clusters 38, 59, and 100:

Rheumatoid_Arthritis.yaml: - "Epigenetic Dysregulation of T Cell Function" with GO:0006338 (chromatin remodeling), GO:0040029 (epigenetic regulation of gene expression), GO:0030217 (T cell differentiation) - Cell types: Th17, CD4+ helper T cell - Downstream link to Autoimmune Response

Systemic_Lupus_Erythematosus.yaml: - "Epigenetic Dysregulation and T Cell Metabolic Reprogramming" with GO:0040029, GO:0030217 - Notes cluster 100 OR=3.7 for SLE (strongest metabolic/stress signal) - Downstream link to Autoantibody Production

These entries enable mechanism-level matching for clusters 38 (MLL2 complex; DOT1L, MEN1), 59 (RNF20, HDAC7), and 100 (ATF4, GLS, CBLB) against dismech pathophysiology GO terms.

2026-02-14 - Project created

• Found existing detailed integration plan at docs/causal-modeling-integration-plan.md
• Supplementary tables S1-S5 already downloaded (S3-S5 were not noted in the plan doc)
• Table S3 contains S-LDSC heritability enrichment (cell type specificity) - important for identifying which traits are well-modeled by K562
• Table S4 appears to be an alternate program annotation (variant analysis)
• Table S5 describes the Perturb-seq datasets used (5 datasets including K562 genome-wide)
• No dedicated GitHub repo for main analysis code exists yet
• T cell follow-up (Zhu/Dann et al. Dec 2025) has a GitHub repo and would extend this to immune traits
• Companion paper (Spence et al.) on gene ranking biases is relevant context
• 12 existing blood/hematological disorders in KB provide reasonable starting coverage
• Key gap: Thalassemia (alpha/beta), Iron Deficiency Anemia, Aplastic Anemia not yet curated