ML-Powered Biomarker Discovery — Design a Publication-Ready Study in 30 Minutes

You: Design a machine learning study to identify diagnostic
biomarkers for early-stage Alzheimer's disease using
publicly available transcriptomic data.

## ML Biomarker Discovery — Early Alzheimer's Disease

### STUDY PATTERN: DEG-to-Diagnostic
(Differential Expression → Feature Selection → ML Model)

### SPECIFIC AIMS
Aim 1: Identify differentially expressed genes between
  early AD patients and cognitively normal controls using
  GEO transcriptomic datasets (DISCOVERY)
Aim 2: Build and validate a diagnostic classifier using
  ensemble ML with cross-validation and external test set
  (MODEL)
Aim 3: Characterize immune cell infiltration patterns
  associated with early AD using computational deconvolution
  (MECHANISM)
Aim 4: Identify hub genes through PPI network analysis and
  validate clinical relevance through survival/progression
  correlation (TRANSLATIONAL)

### DATASET SELECTION
Training datasets (merged):
- GSE63060: Blood transcriptome, 145 AD + 104 controls ✅
- GSE63061: Blood transcriptome, 139 AD + 135 controls ✅
- Combined n=523 (sufficient for ML)

External validation:
- GSE97760: Independent cohort, 36 AD + 36 controls
- GSE122063: Brain tissue (different tissue = stronger
  validation)

Quality checks:
☐ Platform compatibility (all Illumina HumanHT-12)
☐ Batch effect correction (ComBat or limma::removeBatchEffect)
☐ Sample quality (boxplots, PCA for outliers)
☐ Clinical covariate balance (age, sex, APOE status)

### ANALYTICAL PIPELINE (14 Steps)

Step 1: Data preprocessing
  Tool: R/Bioconductor (GEOquery, limma)
  Normalization: quantile normalization
  Batch correction: ComBat (sva package)
  Threshold: Remove probes with detection p>0.05 in
  >50% of samples

Step 2: Differential expression analysis
  Tool: limma
  Model: ~disease_status + age + sex + APOE
  Thresholds: |logFC| > 0.5, adj.p < 0.05
  Expected: 200-800 DEGs

Step 3: Feature selection (3-layer approach)
  Layer 1: DEG filtering (Step 2 output)
  Layer 2: LASSO regression (glmnet, lambda.min)
  Layer 3: Random Forest variable importance (top 50)
  Intersection: genes passing ≥2 of 3 methods
  Expected: 15-40 candidate biomarkers

Step 4: ML model building
  Algorithms: Random Forest, SVM (radial kernel),
  XGBoost, LASSO logistic regression
  Validation: 10-fold cross-validation, repeated 100×
  Metric: AUC-ROC (primary), sensitivity, specificity
  Threshold: AUC > 0.80 for clinical utility

Step 5: Ensemble model
  Method: Weighted average of top 2 performing algorithms
  Weight: proportional to CV-AUC

Step 6: External validation
  Apply model to GSE97760 and GSE122063
  Report: AUC, sensitivity, specificity, calibration plot
  Decision: AUC > 0.75 on external set = proceed to paper

Step 7: CIBERSORT immune deconvolution
  Tool: CIBERSORT (LM22 signature matrix)
  Analysis: Compare immune cell fractions AD vs control
  Correlation: hub genes vs immune cell proportions

[Steps 8-14: PPI network, GO/KEGG enrichment, hub gene
identification, clinical correlation, figure preparation,
manuscript outline]

### WORKLOAD TIERS

| Tier | Timeline | Figures | Target Journal |
|------|----------|---------|---------------|
| Lite | 2-4 weeks | 5-6 | Brief communication |
| Standard | 4-8 weeks | 8-10 | Specialty journal |
| Advanced | 8-14 weeks | 12-16 | High-impact specialty |
| Publication+ | 14+ weeks | 16-20 | High-impact general |

### COMMON PITFALLS (pre-empted)
⚠️ Overfitting: 10-fold CV + external validation prevents this
⚠️ Batch effects: ComBat + PCA visualization before analysis
⚠️ Imbalanced classes: Use AUC not accuracy, apply SMOTE if needed
⚠️ Multiple testing: Benjamini-Hochberg correction for all DEG
⚠️ Data leakage: Feature selection INSIDE CV folds, not before

skills:
  - non-tumor-ml-research-planner
  - medical-research-toolkit
  - pubmed