docs/real_human_data_study_plan.md

# ExtenDRA: Real Human Data Study for Geroprotective Drug Repurposing

## Submission Tracks

This study addresses all three submission categories:

1. **AI and ML-based aging biomarkers** → ExtenDRA proteomic/metabolomic aging clock with uncertainty
2. **AI and ML using genomics, multi-omics, and drug discovery for aging** → Multi-omics integration + perturbation-response drug ranking
3. **AI for clinical translation and trials in aging and aging-related diseases** → Target trial emulation of geroprotective drugs in real EHR + outcomes data

---

## Study Title

**ExtenDRA: Deep Learning-Powered Multi-Omics Aging Clock and Drug Perturbation Framework for Identifying Geroprotective Agents from Human Population Cohorts**

---

## 1. Study Overview

### The Question

> Among FDA-approved medications already taken by thousands of people, which ones slow biological aging — and can we measure this using real human multi-omics data?

### The Design

A three-stage computational study using **real human data only**:

| Stage | What | Data | Output |
|-------|------|------|--------|
| **Stage 1** | Build a multi-omics biological age clock | UK Biobank (53K proteomics + 275K metabolomics + clinical biomarkers) | ExtenDRA-AgeClock: predicted biological age for every participant |
| **Stage 2** | Identify drugs that decelerate biological aging | UK Biobank medication records + GP prescriptions + age acceleration | Ranked list of 30+ drugs/classes associated with reduced age acceleration |
| **Stage 3** | Validate via target trial emulation | UK Biobank longitudinal outcomes (mortality, multimorbidity, frailty) | Causal effect estimates for top geroprotective candidates |

### Why This Works Without Fraud

- UK Biobank is real data from ~500,000 participants with ethical approval
- Medication records are from GP prescriptions (real exposures, not assigned)
- Outcomes (death, disease incidence) are linked from NHS registries
- We use **target trial emulation** (Hernán & Robins framework) to approximate causal inference from observational data — this is an established epidemiological method, not a claim of randomization
- Every result carries uncertainty quantification and sensitivity analyses

---

## 2. Data Sources (All Real Human)

### UK Biobank — Primary Cohort

| Modality | N participants | Variables | Status |
|----------|---------------|-----------|--------|
| **Olink proteomics** | 53,014 | 2,923 proteins (Olink Explore 3072) | Available |
| **Nightingale NMR metabolomics** | 274,371 | 249 metabolites + ratios | Available |
| **Clinical biochemistry** | ~480,000 | 30+ blood/urine biomarkers | Available |
| **Genotyping** | ~488,000 | 800K SNPs + imputed ~90M variants | Available |
| **GP prescriptions** | ~230,000 | Full medication history (BNF codes) | Available |
| **Hospital episode statistics** | ~500,000 | ICD-10 diagnoses, procedures | Available |
| **Death registry** | ~500,000 | Date + cause of death, 15+ years follow-up | Available |
| **Repeat assessment** | ~20,000 | 2nd visit proteomics/metabolomics (4-14yr gap) | Available |

### Medication Classes to Test (>30 drugs, non-cancer focus)

| Category | Drugs/Classes | N exposed (approx UKB) | Aging Rationale |
|----------|---------------|------------------------|-----------------|
| **mTOR-adjacent** | Metformin | ~30,000 | TAME trial pending; reduces IGF-1/mTOR signaling |
| **Statins** | Atorvastatin, Simvastatin, Rosuvastatin | ~100,000 | Anti-inflammatory; associated with lower biological age |
| **ACE inhibitors** | Ramipril, Lisinopril, Enalapril | ~60,000 | RAS system modulates aging; worm/mouse lifespan extension |
| **ARBs** | Losartan, Candesartan, Valsartan | ~30,000 | Same RAS pathway; potentially cleaner mechanism |
| **SGLT2 inhibitors** | Empagliflozin, Dapagliflozin, Canagliflozin | ~5,000+ | Emerging geroprotective: caloric restriction mimetic |
| **GLP-1 agonists** | Semaglutide, Liraglutide, Exenatide | ~3,000+ | Weight loss + anti-inflammatory + emerging longevity signal |
| **Aspirin** | Low-dose aspirin | ~50,000 | Anti-inflammatory; ASPREE trial showed mixed aging results |
| **Beta-blockers** | Bisoprolol, Atenolol, Propranolol | ~40,000 | Heart rate reduction; stress response modulation |
| **Thiazide diuretics** | Bendroflumethiazide, Indapamide | ~30,000 | Blood pressure; some worm longevity data |
| **PPIs** | Omeprazole, Lansoprazole | ~40,000 | Autophagy modulation; controversial aging effects |
| **Allopurinol** | Allopurinol | ~15,000 | Xanthine oxidase inhibitor; reduces oxidative stress |
| **Levothyroxine** | Levothyroxine | ~40,000 | Thyroid hormone; metabolic rate modulation |
| **SSRIs** | Sertraline, Fluoxetine, Citalopram | ~30,000 | Serotonin signaling in C. elegans lifespan |
| **Vitamin D** | Cholecalciferol | ~20,000 | Immune modulation; telomere association |
| **Fibrates** | Fenofibrate, Bezafibrate | ~5,000 | PPARα agonists; metabolic aging |

**Total unique drugs/classes: 35+, all non-cancer primary indication.**

---

## 3. Stage 1: ExtenDRA Multi-Omics Aging Clock

### Architecture

```
Proteins (2,923 Olink) ─► SNN Encoder ─► z_prot (48-dim)
                                                 │
Metabolites (249 NMR) ──► SNN Encoder ─► z_met  │
                                                 ├─► CentralDogmaFusion ─► Predicted Age
Clinical (30 biomarkers) ► SNN Encoder ─► z_clin│
                                                 │
Genotype (PRS for aging) ► SNN Encoder ─► z_gen ┘
```

### Training Target

Chronological age (continuous regression) with:
- L1 loss (robust to outlier ages)
- Trained on 80% of participants, validated on 10%, tested on 10%
- Stratified by sex and ancestry

### Output: Biological Age Gap

```
ΔAge = ExtenDRA_predicted_age − chronological_age
```

- ΔAge > 0: biologically older than expected (accelerated aging)
- ΔAge < 0: biologically younger than expected (decelerated aging)

### Validation (Track 1: AI/ML-based aging biomarkers)

| Validation | Method | Expected |
|------------|--------|----------|
| Mortality prediction | Cox PH: ΔAge → all-cause mortality | HR per SD > 1.3 |
| Disease prediction | Logistic: ΔAge → incident diabetes, CVD, dementia | AUC improvement over chronological age |
| Frailty association | Correlation: ΔAge vs. frailty index | ρ > 0.15 |
| Heritability | GREML on ΔAge | h² ~ 0.10–0.20 |
| Comparison to existing clocks | vs. PhenoAge, GrimAge (methylation), ProtAge (Oh et al.) | Competitive or better |
| Repeat-visit stability | Test-retest on ~20K with 2nd visit | ICC > 0.7 |

---

## 4. Stage 2: Drug Perturbation Scoring (Track 2: Multi-omics + Drug Discovery)

### Design: Medication as Natural Perturbation

Instead of artificial perturbation data, we use **real medication exposure as a natural experiment**.

For each drug class D:
1. Define **exposed group**: participants taking drug D at baseline
2. Define **unexposed group**: matched participants NOT taking D (propensity-score matched on age, sex, BMI, smoking, comorbidities, other medications, deprivation, ethnicity)
3. Compute:

```
ΔAge_exposed = mean biological age gap among exposed
ΔAge_unexposed = mean biological age gap among matched unexposed
Drug_aging_effect = ΔAge_exposed − ΔAge_unexposed
```

A negative Drug_aging_effect means the drug is associated with younger biological age.

### ExtenDRA-Perturb Integration

Use the perturbation module to:
1. **Encode drug**: SMILES → Morgan fingerprint / ChemBERTa embedding
2. **Encode baseline state**: participant multi-omics → ExtenDRA latent state
3. **Predict perturbation response**: FiLM conditioning → predicted proteomic/metabolomic shift
4. **Score**: does the predicted shift oppose the aging direction?

This connects the heuristic drug-ranking to a learned perturbation model trained on real exposed/unexposed proteomic differences.

### Confounding Control

| Confounder | Method |
|------------|--------|
| Indication bias (sicker people take drugs) | Propensity score matching |
| Immortal time bias | Landmark analysis (require 1yr exposure before clock measurement) |
| Polypharmacy | Adjust for total medication count + specific co-medications |
| Reverse causation | Sensitivity: exclude participants with prevalent disease at baseline |
| Genetic confounding | Negative control outcomes + MR validation |

### Mendelian Randomization Validation

For drugs with known protein targets:
1. Find pQTLs (protein quantitative trait loci) for the target protein in UKB
2. Use cis-pQTL as genetic instrument
3. MR outcome: ΔAge (biological age acceleration)
4. If MR confirms same direction as observational → stronger causal evidence

Example: Statins target HMGCR → HMGCR pQTL → test causal effect on ΔAge.

---

## 5. Stage 3: Target Trial Emulation (Track 3: Clinical Translation)

### Framework

For top 10 geroprotective candidates from Stage 2, run formal **target trial emulations** (Hernán & Robins 2016):

| Trial Element | Emulation |
|---------------|-----------|
| Eligibility | Age 50–70, no prior cancer/CVD/dementia at index date |
| Treatment strategies | Initiate drug D vs. do not initiate |
| Assignment | Clone-censor-weight (inverse probability of censoring weights) |
| Follow-up start | Date of eligibility + 1 year grace period |
| Outcomes | All-cause mortality, incident multimorbidity (≥3 chronic diseases), frailty transition, composite aging endpoint |
| Follow-up | Up to 15 years (UKB follow-up) |
| Analysis | Pooled logistic regression with stabilized IPCW |

### Outcomes (Clinical Translation)

| Outcome | Definition | Follow-up |
|---------|-----------|-----------|
| All-cause mortality | NHS death registry | 15 years |
| Cardiovascular events | ICD-10 I20-I25, I60-I69 | 15 years |
| Type 2 diabetes | E11 new diagnosis | 15 years |
| Dementia | F00-F03, G30 new diagnosis | 15 years |
| Multimorbidity transition | 0-2 → 3+ chronic diseases | 15 years |
| Frailty transition | Fried criteria equivalent | At repeat visit |
| Composite aging endpoint | First of: death, dementia, frailty, or multimorbidity | 15 years |

### Sensitivity Analyses

1. Active comparator (e.g., metformin vs. sulfonylurea in diabetics)
2. Dose-response (higher dose → more deceleration?)
3. Duration-response (longer exposure → more benefit?)
4. E-value for unmeasured confounding
5. Negative control exposures (drugs with no plausible aging mechanism)
6. Negative control outcomes (outcomes drug shouldn't affect)

---

## 6. ExtenDRA Platform Demonstration

### What ExtenDRA Adds Over Existing Methods

| Capability | Standard approach | ExtenDRA |
|-----------|-------------------|----------|
| Aging biomarker | Single-omic (methylation OR proteomics) | Multi-omic fusion following central dogma |
| Drug ranking | Literature-based or single-omic association | Perturbation-conditioned response prediction with uncertainty |
| Interpretation | Black-box age prediction | Per-protein/pathway attribution via Integrated Gradients |
| Uncertainty | Point estimates only | Ensemble + heteroscedastic variance → per-drug confidence intervals |
| Cross-validation | Single cohort | UKB discovery → longitudinal replication → MR validation |

### Platform Components Used

```
Stage 1 (Clock):
  extendra/models.py    → CentralDogmaFusion, OmicsEncoder, SNNStack
  extendra/losses.py    → L1 age regression + regularization
  extendra/causal.py    → Integrated Gradients for protein attribution

Stage 2 (Drug Perturbation):
  extendra/perturb/encoder.py      → Drug SMILES → embedding
  extendra/perturb/conditioning.py → FiLM: drug embedding modulates participant state
  extendra/perturb/model.py        → ExtenDRAPerturb: predict proteomic shift under drug
  extendra/perturb/ensemble.py     → Uncertainty quantification
  extendra/perturb/evaluate.py     → Pearson-Δ, DES@K, direction-match
  extendra/perturb/attribution.py  → Pathway-level drug mechanism

Stage 3 (Clinical Translation):
  extendra/pipeline.py  → End-to-end: data → clock → drug score → outcome validation
  extendra/trainer.py   → Training with Trackio monitoring
```

---

## 7. Expected Results and Deliverables

### Primary Deliverables

1. **ExtenDRA-AgeClock** — a multi-omics biological age predictor:
   - Trained on 53K UKB participants with proteomics + metabolomics
   - MAE < 4 years for chronological age
   - HR > 1.3 per SD of age acceleration for mortality
   - Publicly deposited model weights

2. **Geroprotective Drug Ranking** — 35+ medications scored by age-deceleration effect:
   - Each drug: effect size + 95% CI + pathway attribution + MR p-value
   - Expected top hits: metformin, SGLT2i, GLP-1 agonists, ACE inhibitors
   - Novel candidates: drugs with unexpected age-deceleration signals

3. **Target Trial Results** — top 10 drugs with formal causal estimates:
   - Hazard ratios for composite aging endpoint
   - Number needed to treat (NNT) for 5-year aging prevention
   - Dose-response and duration-response curves

4. **Open-source ExtenDRA platform** — reusable for other cohorts:
   - Code: https://huggingface.co/vedatonuryilmaz/ExtenDRA-Longevity
   - Pre-trained clock weights (shareable under UKB data return agreement)
   - Perturbation scoring module usable on any cohort with proteomics + medication data

### Expected Positive Controls (Should Replicate)

| Drug | Expected direction | Literature support |
|------|-------------------|-------------------|
| Metformin | Age deceleration | TAME trial rationale; diabetic survival paradox |
| Statins | Age deceleration | Anti-inflammatory; multiple observational studies |
| ACE inhibitors | Age deceleration | RAS modulation; worm/mouse lifespan |
| Rapamycin analogs | Unknown in humans | Strong pre-clinical; no UKB exposure data |

### Expected Negative Controls (Should Show No Effect)

| Drug | Expected | Rationale |
|------|----------|-----------|
| Antihistamines (cetirizine) | No effect | No aging mechanism |
| Antacids (calcium carbonate) | No effect | No aging mechanism |
| Topical treatments | No effect | No systemic exposure |

---

## 8. Timeline

| Month | Milestone |
|-------|-----------|
| 1–2 | UKB data access + QC + cohort assembly |
| 2–3 | Train ExtenDRA-AgeClock on proteomics + metabolomics |
| 3–4 | Validate clock against mortality, disease, frailty |
| 4–5 | Compute drug-specific age acceleration effects (35+ drugs) |
| 5–6 | Train ExtenDRA-Perturb on exposed/unexposed proteomic differences |
| 6–8 | Target trial emulations for top 10 candidates |
| 8–9 | MR validation for drugs with pQTL instruments |
| 9–10 | Sensitivity analyses + negative controls |
| 10–11 | Manuscript + poster preparation |
| 11–12 | Submission + code/model release |

---

## 9. Why Funders Should Care

### Market

- Global longevity market: $44B in 2024, projected $93B by 2030 (CAGR 13%)
- Drug repurposing: 10× cheaper than de novo discovery ($300M vs $2.6B)
- UK Biobank is the world's largest population cohort with deep molecular phenotyping

### Differentiation

- Not another methylation clock — **first multi-omic perturbation-response aging framework** that directly scores drugs
- Not just association — **causal inference** via target trial emulation + MR
- Not black-box — **interpretable per-protein/pathway attribution** that pharma partners can act on
- **Immediate clinical translation**: all drugs are already FDA-approved, already taken by thousands

### IP/Moat

- The trained ExtenDRA-AgeClock on UKB proteomics is a proprietary asset
- Drug ranking + causal evidence = partnership leverage with pharma
- Platform generalizes to other cohorts (All of Us, ARIC, Framingham)

### Risk Mitigation

- UKB data access is standard (approved in weeks, not months)
- All drugs are real, all outcomes are real, all methods are published
- Target trial emulation is accepted by FDA for regulatory evidence
- Negative controls prevent false discovery claims
- Every result carries uncertainty intervals

---

## 10. Ethical Considerations

- UK Biobank data used under approved application (ethics: North West Multi-Centre Research Ethics Committee, 16/NW/0274)
- No individual-level results returned to participants
- Drug repurposing candidates are identified computationally — no clinical recommendations made without Phase III trials
- All code and non-restricted model artifacts will be open-sourced
- Equity: analysis includes ancestry-stratified results (UKB has ~5% non-European ancestry; limitations acknowledged)

---

## 11. Summary: Three-Track Alignment

| Track | ExtenDRA Contribution |
|-------|----------------------|
| **AI/ML aging biomarkers** | Multi-omics biological age clock with uncertainty, validated against mortality/disease/frailty |
| **AI/ML genomics + multi-omics + drug discovery** | Perturbation-response model trained on real medication-exposed proteomic shifts; MR-validated drug targets |
| **AI for clinical translation** | Target trial emulation of 10+ drugs with HR estimates, NNT, dose-response — ready for Phase IV trial prioritization |

---

## References (Key Published Precedents)

1. Oh et al. (2023) "Organ aging signatures in plasma proteomics track health and disease." Nature 624, 164–172. [53K UKB Olink; organ-specific aging clocks predict mortality]
2. Argentieri et al. (2024) "Proteomic aging clock predicts mortality and identifies targets for healthy longevity." Nature Medicine. [UKB proteomic age → mortality HR 1.67/SD]
3. Timmers et al. (2020) "Multivariate genomic scan implicates novel loci and haem metabolism in human ageing." Nature Communications 11, 3570. [pQTL + MR for aging/longevity]
4. Lehallier et al. (2019) "Undulating changes in human plasma proteome profiles across the lifespan." Nature Medicine 25, 1843–1850. [2,925 proteins × 4,263 people; aging waves]
5. Hernán & Robins (2016) "Using big data to emulate a target trial when a randomized trial is not available." American J Epidemiol 183(8):758–764.
6. Sun et al. (2023) "Plasma proteomics identifies genetic associations with disease risk and drug targets." Nature Genetics 55, 1599–1608. [3,000+ pQTLs for MR drug target validation]
7. Julkunen et al. (2023) "Atlas of plasma NMR biomarkers for health and disease in 118,461 individuals." Nature Communications 14, 604. [Nightingale metabolomics aging associations]
8. TAME Trial Investigators (2024) "Targeting Aging with Metformin (TAME): rationale and design." [The first anti-aging RCT; metformin is the benchmark]