Problem Framing

What is “Resilience” in This Project?

In the context of this research, resilience refers to the ability of Crassostrea virginica (Eastern oyster) to tolerate or resist stress from Perkinsus marinus (Dermo disease) infection and other environmental challenges.

Defining the Phenotype

Throughout this project, phenotype definition evolved significantly:

!!! warning “Big Lesson #2: Traits Were Oversimplified” Early attempts used generalized trait definitions across studies (e.g., “stress” vs. “control”). This oversimplification failed because:

- Different studies measured different stressors (disease, temperature, salinity)
- Trait effects were weaker than study-specific effects
- Batch correction couldn't compensate for fundamental differences in experimental design

Current approach: Focus on specific, comparable phenotypes:

  • Tolerant/Resistant: Oysters that survived or showed low infection intensity
  • Sensitive/Susceptible: Oysters that died or showed high infection intensity
  • Control vs. Treated: Within-study comparisons before examining resilience

The Datasets

This project integrates multiple RNA-seq datasets from C. virginica exposed to P. marinus:

Primary Datasets

  • Dataset 1: [Description of dataset 1 characteristics]
  • Dataset 4: Injected group samples
  • Dataset 5: [Description of dataset 5 characteristics]
  • Additional datasets analyzed for specific questions

Data Characteristics

Common challenges:

  • Batch effects: Study-specific effects consistently stronger than trait effects
  • Sample size limitations: Variable n across studies
  • Time point variation: Sampling at different post-exposure times
  • Technology differences: Including TAG-seq vs. standard RNA-seq (issue #26, #28)

Key Constraints & Design Decisions

1. Integration vs. Post-Integration Analysis

!!! example “Big Lesson #1: When Integrated Analysis Fails” Integrated data analysis does not work when data is noisy (too much within and/or across study variation) and signal is not strong enough.

**Initial approach (Failed):**
- Pooled all datasets together
- Attempted batch correction (COMBAT, RemoveBatchEffect)
- Expected to see trait-based clustering

**Outcome:** Study-specific effects dominated; trait separation was poor

**Pivot:** Adopted post-data integration approach:
1. Analyze each dataset independently
2. Compare results across datasets
3. Identify reproducible signatures

2. Fixed vs. Random Effects

Early attempts didn’t properly account for:

  • Sample size differences across studies
  • Control group considerations
  • Study as a random effect in mixed models

3. Normalization Timing

!!! info “When Normalization Happens” In the nf-core/differentialabundance pipeline:

- PCAs are generated **before** differential abundance analysis
- Normalization (VST, TPM) happens during analysis
- Batch correction (if applied) occurs on normalized counts

Understanding this order is critical for interpreting preliminary QC plots

4. Innate vs. Reactive Biomarkers

!!! success “Big Lesson #3: Don’t Filter Out Innate Signals” Biomarkers may exist in control groups if they represent innate resilience traits (genes that are constitutively different in resistant vs. sensitive individuals).

**Implication:** The stepwise approach (filter controls first) may inadvertently remove true biomarkers

**Resolution:** Developed alternative classifier approach that preserves innate signals (see [Two-Step Classifier](/field-guide-src/docs/pipelines/classifier-path/))

See issue #53 and notebook post for full innate vs. reactive analysis.

Biological Context

Perkinsus marinus (Dermo)

  • Major pathogen in oyster aquaculture
  • Causes dermo disease (mortality and reduced growth)
  • Variable infection response across oyster populations
  • Understanding resistance is key to breeding programs

Molecular Signatures of Resilience

This project aims to identify gene expression patterns that:

  1. Predict tolerance/resistance phenotypes
  2. Are reproducible across independent studies
  3. Are biologically interpretable (pathway-informed)
  4. Could be assayable in breeding programs (minimal gene panels)

Research Questions Evolution

Initial Questions (December 2024)

  • Can we identify shared stress response genes across multiple studies?
  • Do RNA-seq datasets cluster by trait when integrated?

Refined Questions (January-April 2025)

  • How do batch effects and study design differences limit integration?
  • Can batch correction methods recover trait-based signal?
  • Which normalization approaches work best for meta-analysis?

Current Questions (August-September 2025)

  • What is the minimal gene set that distinguishes tolerant from sensitive oysters?
  • Are biomarkers innate or reactive (expressed before vs. after stress)?
  • Can classifiers trained on one study predict phenotypes in independent studies?

Success Criteria

A successful biomarker panel should:

✅ Show reproducible differential expression across datasets
✅ Maintain predictive accuracy in Leave-One-Study-Out (LOSO) validation
✅ Be small enough for practical assay development (< 10 genes)
✅ Have biological interpretability (annotatable, pathway-linked)
✅ Distinguish phenotypes in both control and treated conditions

Next: Explore validated analysis approaches in Pipelines