Course Project: Your Own Exposome-Wide Association Study

Overview

In this project, you will conduct an original ExWAS by selecting or deriving a phenotype of your choice and systematically testing its association with hundreds of environmental exposures using the nhanespewas package. The goal is not just to run the pipeline — it is to think critically about what you find, what it means, and what it doesn’t mean.

Part 1: Phenotype Selection or Derivation (10%)

Choose one of the following approaches:

Option A: Select an existing phenotype

Pick a continuous or binary phenotype from the p_catalog in nhanespewas that has not been extensively studied in the ExWAS literature (i.e., not BMI, glucose, or HbA1c). Justify your choice: why is this phenotype interesting from a public health or biological perspective?

Option B: Derive a novel phenotype

Construct a composite or derived phenotype from NHANES variables. Examples:

  • Metabolic syndrome score: combine waist circumference, triglycerides, HDL, fasting glucose, and blood pressure into a continuous risk score
  • Kidney function index: derive eGFR from serum creatinine using the CKD-EPI equation
  • Inflammation composite: combine CRP, white blood cell count, and fibrinogen
  • Cardiometabolic ratio: triglyceride-to-HDL ratio (a marker of insulin resistance)
  • Pulmonary function ratio: FEV1/FVC ratio
  • Liver function composite: combine ALT, AST, GGT into a score
  • Allostatic load: a multi-system physiological dysregulation index

You may also define a binary phenotype by thresholding a continuous variable (e.g., eGFR < 60 for chronic kidney disease, metabolic syndrome present/absent by ATP-III criteria).

Deliverable: A 1-paragraph justification of your phenotype, including its clinical or biological relevance, how it was measured or derived, and any limitations of the definition.

Part 2: Conduct the ExWAS (20%)

Using the tools from Modules 4-5, run a full ExWAS:

  1. Connect to the nhanespewas database
  2. Run pe_flex_adjust() (or logistic_e_flex_adjust() for binary phenotypes) across all available exposures
  3. Use at least two adjustment models (e.g., unadjusted and fully adjusted) to assess sensitivity
  4. Apply Benjamini-Hochberg FDR correction
  5. Report: number of exposures tested, number significant at FDR < 0.05, number significant at Bonferroni threshold

You may use AI tools (e.g., Claude Code) to help write the pipeline code. If you do, cite how you used it.

Deliverable: Commented R code and a summary table of results.

Part 3: Visualize and Describe Results (20%)

Create the following visualizations:

  1. Volcano plot: effect size vs. -log10(p-value), with FDR and Bonferroni thresholds marked, top hits labeled
  2. Effect size comparison: plot estimates from the unadjusted model vs. the fully adjusted model — highlight any sign reversals
  3. One additional visualization of your choice (e.g., delta R-squared bar chart, exposure category breakdown, forest plot for a top hit across NHANES waves)

For each plot, write a 2-3 sentence interpretation of what the plot shows.

Deliverable: Three figures with captions and interpretations.

Part 4: Interpret Your Top 5 Hits (30%)

For each of your top 5 FDR-significant associations, write a structured interpretation that addresses:

a) Biological plausibility

  • Is there a known biological mechanism linking this exposure to your phenotype?
  • Cite at least one relevant paper or review per hit (a brief literature search is expected)

b) Confounding assessment

  • Draw a DAG specific to this exposure-phenotype pair
  • Identify at least one confounder that the standard adjustment set (age, sex, race/ethnicity, income) likely misses
  • Predict whether your estimate is likely biased toward or away from the null, and in which direction

c) Sensitivity to adjustment

  • Did the estimate change substantially between the unadjusted and adjusted models?
  • If yes, what does this imply about confounding? If no, does stability mean the estimate is unconfounded?

d) Replication potential

  • Would you expect this finding to replicate in an independent NHANES wave? Why or why not?
  • What would strengthen or weaken your confidence in this association?

Deliverable: A structured write-up (approximately 1 page per hit, 5 pages total).

Part 5: Limitations and Future Directions (20%)

Write a 2-page discussion addressing:

  1. The confounding problem: How does the per-association confounding issue (Module 3) affect your specific results? Which of your top hits are most vulnerable to unmeasured confounding, and why?

  2. Cross-sectional limitations: For which of your hits is reverse causation a plausible concern? Give a specific example of how the phenotype could affect the exposure rather than vice versa.

  3. Measurement considerations: Are any of your top exposures measured with known limitations (short half-life, high assay variability, below-detection-limit issues)? How might measurement error affect your estimates?

  4. A validation plan: If you had unlimited resources, how would you validate your most promising finding? Describe a triangulation strategy (at least 3 approaches from Module 7) tailored to your specific hit.

  5. What would LC-HRMS add?: Would untargeted exposomics (Module 10) be useful for your phenotype? What unknown exposures might be relevant that NHANES does not measure?

Deliverable: A 2-page discussion.

Grading Rubric

Component Weight Key criteria
Phenotype selection/derivation 10% Originality, justification, appropriate definition
ExWAS execution 20% Correct implementation, multiple adjustment models, FDR
Visualization 20% Clear figures, appropriate plot choices, correct interpretation
Top 5 interpretation 30% DAG quality, confounding reasoning, biological plausibility, literature engagement
Limitations discussion 20% Depth of reasoning, specificity to own results, validation plan quality

Important Notes

  • Code can be automated; interpretation cannot. The majority of your grade (70%) comes from interpretation, reasoning, and discussion — not from running the pipeline.
  • There are no “wrong” phenotypes. A null ExWAS (few or no significant hits) is a valid result if interpreted thoughtfully. Discuss why you found few hits — is it low power, weak effects, or a phenotype truly unrelated to the measured exposome?
  • Honest assessment of limitations is valued. Acknowledging what your analysis cannot tell you is as important as reporting what it can.
  • Use of AI tools is permitted but must be cited. If you use Claude Code, ChatGPT, or similar tools to generate code or draft text, state how you used them and what you modified. The interpretation must be your own.