Plan: DisMech as Validation Resource for Causal Gene-to-Trait Analyses
The Idea
Given outputs from causal modeling pipelines (GWAS + Perturb-seq → gene programs → traits), can DisMech's curated pathophysiology knowledge explain, validate, or contextualize the discovered relationships?
This positions DisMech as a benchmark/interpretation layer rather than a data store.
Background: The Ota et al. Methodology
The paper "Causal modelling of gene effects from regulators to programs to traits" (Ota et al., Nature 2025) builds causal graphs:
Gene (regulator) --[β]--> Program --[effect]--> Trait/Phenotype
↑
Gene (member) ----[γ]----/
Key data types: - γ (gamma): Gene effect on trait from LoF burden tests (UK Biobank, 454K participants) - β (beta): Regulatory effect from Perturb-seq knockdowns (K562 cells, 9,498 genes) - Programs: Co-expressed gene modules (60 identified via consensus NMF)
Key outputs: - Program-trait effect sizes with directions - Gene-program regulatory relationships - Causal chains explaining how genes affect traits via biological programs - 73% accuracy in predicting effect directions for top GWAS hits
Use Cases for DisMech
1. Validate Discovered Gene-Program-Trait Relationships
Question: Does a computationally discovered pathway match curated knowledge?
| Pipeline Output | DisMech Query |
|---|---|
| GATA1 → erythroid program → MCH | Do blood disorders document GATA1 affecting hemoglobin via erythropoiesis? |
| BCL2 → apoptosis program → lymphocyte count | Do immune disorders link BCL2 to apoptosis affecting lymphocytes? |
Validation levels: - CONFIRMED: Gene, process, and phenotype all documented with evidence - PARTIAL: Some elements present but incomplete chain - NOVEL: Not in DisMech (curation candidate) - CONTRADICTED: DisMech documents opposite effect
2. Explain Programs via Pathophysiology
Question: What does "Program 17" actually mean in disease context?
Pipeline outputs are anonymous gene modules (Program 1, Program 2, ...). DisMech can provide clinical/mechanistic context:
| Program GO Annotation | DisMech Diseases | Clinical Context |
|---|---|---|
| GO:0007049 (cell cycle) | CML, Retinoblastoma | "Uncontrolled proliferation drives tumor growth" |
| GO:0006915 (apoptosis) | SLE, Autoimmune diseases | "Defective clearance of apoptotic cells triggers autoimmunity" |
| GO:0030218 (erythrocyte differentiation) | Sickle Cell, Thalassemia | "Ineffective erythropoiesis causes anemia" |
3. Prioritize Novel Findings
Question: Which discovered relationships are genuinely novel vs. already known?
Given 500 gene-trait associations from a pipeline, DisMech can partition into: - Known: Already documented (low priority for follow-up) - Novel: Not in literature (high priority for experimental validation) - Contradicted: Conflicts with literature (needs investigation)
4. Identify Mechanistic Gaps in DisMech
Question: Where does computational analysis find things missing from curated knowledge?
If pipeline discovers "Gene X → autophagy → Trait Y" but DisMech's Disease Y entry lacks autophagy in pathophysiology, this flags a curation gap.
Feedback loop: Pipeline discoveries improve DisMech coverage.
5. Cross-Disease Program Analysis
Question: Do different diseases share causal programs?
| Shared Program | Diseases | Implication |
|---|---|---|
| Inflammation (GO:0006954) | RA, Crohn's, Psoriasis | Common therapeutic targets |
| Fibrosis (GO:0030198) | IPF, Liver Cirrhosis, Systemic Sclerosis | Shared mechanism despite different organs |
Data Requirements
From Ota et al. Pipeline
- Program definitions: Gene lists per program + GO annotations
- Program-trait effects: Effect sizes and directions for each program-trait pair
- Gene-program relationships: Which genes regulate which programs (β values)
- Gene-trait effects: Direct LoF burden test results (γ values)
Note: Check if supplementary tables from paper are publicly available.
From DisMech
- Gene mentions: Genes in
geneticandpathophysiology.genessections - Process annotations: GO terms in
pathophysiology.biological_processes - Phenotype annotations: HP terms in
phenotypes - Cell type context: CL terms in
pathophysiology.cell_types - Evidence: PMIDs supporting each relationship
Mapping Requirements
| Pipeline Entity | DisMech Entity | Mapping Strategy |
|---|---|---|
| Gene symbol | GeneDescriptor | Direct match or HGNC lookup |
| Program GO term | BiologicalProcessTerm | Exact or ancestor match via GO hierarchy |
| Trait | Phenotype HP term | Manual curation of trait-to-HPO mappings |
| Cell type | CellTypeTerm | Match K562 → CL:0000255 (erythroid) |
Evaluation Criteria
Quantitative Metrics
- Coverage: % of pipeline gene-trait pairs that DisMech can evaluate
- Confirmation rate: % of top pipeline hits confirmed by DisMech
- Novelty rate: % of pipeline discoveries not in DisMech
- Contradiction rate: % of pipeline results conflicting with DisMech
Qualitative Assessment
- Do confirmed findings have strong evidence (experimental PMIDs)?
- Are novel findings biologically plausible given DisMech context?
- Can contradictions be resolved by examining evidence quality?
DisMech Gaps to Address
Current Strengths
- 55+ disorders with curated pathophysiology
- GO/HP/CL term bindings enabling semantic matching
- Evidence with PMIDs for traceability
- Structured pathophysiology with biological processes and cell types
Gaps
| Gap | Impact | Remediation |
|---|---|---|
Inconsistent modifier usage |
Cannot validate effect directions | Curation pass to add INCREASED/DECREASED |
| Genes often missing from pathophysiology | Low gene coverage | Add genes to pathophysiology entries |
| No explicit gene→process links | Cannot trace causal chains | Add gene slot to BiologicalProcessDescriptor |
| Blood disorders underrepresented | Poor coverage for Ota et al. test case | Prioritize Thalassemia, Polycythemia, etc. |
Pilot: Blood Traits
The Ota et al. paper focuses on three blood traits: - MCH (mean corpuscular hemoglobin) - RDW (red cell distribution width) - IRF (immature reticulocyte fraction)
Relevant DisMech Disorders
| Disorder | Relevance | Current Coverage |
|---|---|---|
| Sickle Cell Disease | Hemoglobin, erythrocytes | Good - has GO terms, cell types, evidence |
| Autoimmune Hemolytic Anemia | Hemolysis, RBC destruction | Moderate |
| Fanconi Anemia | Bone marrow failure | Moderate |
| G6PD Deficiency | RBC fragility | Check coverage |
| Hemochromatosis | Iron metabolism | Check coverage |
Missing Disorders to Add
- Thalassemia (alpha and beta)
- Polycythemia vera
- Iron deficiency anemia
- Aplastic anemia
- Myelodysplastic syndromes
Output Formats
1. Validation Report (per gene-program-trait)
gene: GATA1
program: erythroid_differentiation
program_go: GO:0030218
trait: mean_corpuscular_hemoglobin
effect_direction: positive
validation:
status: CONFIRMED
matching_disorders:
- disease: Sickle Cell Disease
pathophysiology_entry: "Red Blood Cell Sickling"
process_match: GO:0030218 (erythrocyte differentiation)
gene_mentioned: false # gap identified
evidence: PMID:24277079
confidence: MEDIUM
notes: "Process confirmed, gene not explicitly mentioned"
curation_action: "Add GATA1 to SCD pathophysiology"
2. Summary Statistics
Total pipeline relationships evaluated: 180
- CONFIRMED: 73 (41%)
- PARTIAL: 45 (25%)
- NOVEL: 52 (29%)
- CONTRADICTED: 10 (6%)
Coverage by program type:
- Cell cycle programs: 85% evaluable
- Erythroid programs: 92% evaluable
- Autophagy programs: 34% evaluable (gap)
3. Curation Candidates
Ranked list of novel findings with high effect sizes that warrant adding to DisMech.
4. Gap Report
Which GO terms from programs lack DisMech coverage, prioritized by frequency in pipeline results.
Implementation Phases
Phase 1: Data Assembly
- Obtain Ota et al. supplementary data
- Export DisMech blood disorders to queryable format
- Create trait-to-HPO mapping for MCH, RDW, IRF
Phase 2: Proof of Concept
- Manual validation of 20 top gene-program-trait relationships
- Document matching logic and edge cases
- Assess DisMech coverage and gaps
Phase 3: Systematic Evaluation
- Automate matching across all pipeline outputs
- Generate validation report
- Quantify confirmation/novelty/contradiction rates
Phase 4: Bidirectional Improvement
- Use novel findings to improve DisMech coverage
- Re-run validation to measure improvement
- Publish methodology and results
Open Questions
-
Trait mapping: How to map blood lab values (MCH) to HPO terms? Direct terms exist (HP:0025066 "Decreased mean corpuscular hemoglobin") but may need expert review.
-
Cell type specificity: Ota et al. used K562 (erythroleukemia line). How generalizable are programs to other cell types in DisMech pathophysiology?
-
Effect direction validation: DisMech has
modifierbut it's underused. How much curation effort to systematically add directions? -
Semantic matching depth: Should GO term matching use exact match, or traverse hierarchy (e.g., "erythrocyte differentiation" matches "hematopoiesis")?
-
Evidence quality weighting: Should we weight DisMech evidence by experimental type (IDA > TAS > NAS)?
References
- Ota et al. Nature 2025 - Main methodology
- bioRxiv preprint - Open access version
- PMC full text - Full text with supplements
Assets
- Paper PDF:
docs/assets/Causal modelling of gene effects from regulators to programs to traits.pdf - Table S1:
docs/assets/Ota_2025_TableS1.xlsx- Program annotations (60 programs) - Curated annotations (e.g., "Hemoglobin synthesis", "Cell cycle (S phase)", "Autophagosome")
- Representative GO terms
- Top 10 genes per program
- Representative transcription factors
- Table S2:
docs/assets/Ota_2025_TableS2.xlsx- Trait definitions (54 traits) - RBC traits: MCH, MCV, RDW, reticulocyte count, hemoglobin, etc.
- Other blood traits: platelet count, WBC, lymphocyte count, etc.
- Serum biomarkers: albumin, cholesterol, glucose, bilirubin, etc.
- Anthropometric: BMI, height, bone mineral density
Key Programs for Blood Trait Validation
| Program | Annotation | GO Term | Relevance to DisMech |
|---|---|---|---|
| P4 | Cell cycle (S phase, DNA replication) | DNA replication | CML, Retinoblastoma |
| P6 | Cell cycle (G2M checkpoint) | G2M checkpoint | Cancers |
| P16 | Autophagosome | autophagosome | Neurodegeneration, inflammation |
| P27 | Platelet activation (Mega, IKZF1) | platelet activation | Immune thrombocytopenia |
| P28 | RBC (glycoproteins) | HEME_METABOLISM | Sickle Cell, Thalassemia |
| P35 | TNF signaling (stress response) | TNFA_SIGNALING_VIA_NFKB | Autoimmune diseases |
| P40 | Hemoglobin synthesis | HEME_METABOLISM | Sickle Cell, Thalassemia, Anemia |
| P45 | TNF signaling (apoptosis) | TNFA_SIGNALING_VIA_NFKB | SLE, RA |
Data Not Yet Available
The bioRxiv preprint only includes Tables S1 and S2. Additional data mentioned in the paper: - Table S3: Program-trait effect sizes (needed for quantitative validation) - Gene-level burden test results (γ values) - GitHub repository with analysis code (not yet public)
These may become available with the Nature publication or upon request to authors.