Chapter 18: Integrated Clinical Proteomics: The Full Analytical Pipeline¶

18.1 Overview¶

This chapter details the comprehensive computational workflow used to identify proteomic signatures of Alzheimer's Disease (AD) by integrating data from the ADNI and ADRC cohorts. The analysis proceeds sequentially from raw data cleaning to systems-level network analysis (WGCNA), functional enrichment, and machine learning.

18.2 Data Preprocessing and Harmonization¶

Scripts: 00_Data_preprocessing.Rmd, 02a_ADNI_ADRC_harmonized.Rmd, 02b_PCA_full.Rmd

The initial phase focused on ensuring data quality and comparability between cohorts.

Quality Control and Cleaning¶

Raw proteomic data often contains missing values and technical artifacts. * Missingness Filtering: Proteins with high missingness (e.g., >20%) across samples were removed to ensure robust downstream analysis. * Sample Filtering: Samples with incomplete core clinical metadata (Age, Sex, Diagnosis) were excluded.

Cross-Cohort Harmonization¶

To integrate the ADNI and ADRC datasets, which may have been processed at different times or sites, we applied harmonization techniques to mitigate batch effects. * Batch Correction: Statistical methods were used to align the expression distributions of shared proteins between the two cohorts, ensuring that biological signals were preserved while technical differences were minimized.

Exploratory Data Analysis (PCA)¶

Principal Component Analysis (PCA) was performed on the harmonized data. * Variance Explained: We assessed the proportion of variance captured by the top principal components. * Visualization: PCA plots allowed us to visualize the separation between cohorts before and after harmonization, as well as to identify and remove potential outliers.

18.3 Network Construction (WGCNA)¶

Script: 03_WGCNA_running.Rmd

Illustration of WGCNA Modules

We employed Weighted Gene Co-expression Network Analysis (WGCNA) to identify clusters (modules) of co-expressed proteins. This systems biology approach moves beyond single-protein analysis to identify functional units.

Soft Thresholding¶

A key feature of WGCNA is soft thresholding. * Scale-Free Topology: We analyzed the scale-free topology fit index for various powers (\(\beta\)). * Power Selection: We selected the lowest power \(\beta\) that resulted in a high scale-free topology fit (\(R^2 > 0.85\) or similar), ensuring the network reflects biological reality (few hubs, many peripheral nodes). * Adjacency Matrix: The correlation matrix was raised to the power \(\beta\) (\(|correlation|^\beta\)) to create a weighted adjacency matrix, suppressing weak correlations and emphasizing strong ones.

Module Detection¶

Topological Overlap Matrix (TOM): The adjacency matrix was converted into a TOM to measure the interconnectedness of proteins.
Hierarchical Clustering: We performed average linkage hierarchical clustering on the dissimilarity matrix (1-TOM).
Dynamic Tree Cut: Modules were identified using the dynamic tree cut algorithm, which is capable of detecting nested clusters.

Eigengene Calculation¶

For each module, we calculated the Module Eigengene (ME). The ME is the first principal component of the module's expression matrix and serves as a synthetic representative profile for the entire module.

18.4 Functional Annotation and Enrichment¶

Scripts: 04_Functional_Analysis.Rmd, 07a_Functional_Analysis_GSEA.Rmd, 07b_Functional_Analysis_ORA.Rmd

To understand the biological significance of the identified modules, we performed enrichment analyses.

Over-Representation Analysis (ORA)¶

Method: We used hypergeometric testing to determine if specific Gene Ontology (GO) terms or KEGG pathways were present in a module more often than expected by chance.
Databases: GO (Biological Process, Molecular Function, Cellular Component), Reactome, and KEGG.

Gene Set Enrichment Analysis (GSEA)¶

Ranked Lists: Proteins were ranked based on their correlation with clinical traits or module membership.
Enrichment: We performed GSEA to identify coordinated pathway alterations that might not be detected by ORA alone, as it considers the entire ranked list rather than a fixed threshold.

18.5 Module Scoring and Feature Selection¶

Script: 05_Module_Scoring_Selection.Rmd

We evaluated the predictive power of the identified modules and selected key features for modeling.

Module Scores: Composite scores were calculated for each module (e.g., average expression or eigengene value) to relate module activity to clinical traits.
Hub Identification: We calculated Module Membership (kME), which is the correlation between a protein's expression and the module eigengene. Proteins with high kME are considered "hubs" and are likely drivers of the module's biological function.

18.6 Machine Learning for Biomarker Discovery¶

Script: 06_ADNI_ADRC_ML.Rmd

We trained machine learning models to predict AD status and clinical outcomes using the identified proteomic signatures.

Model: Elastic Net Regularized Regression¶

We utilized the Elastic Net algorithm (glmnet package), which combines L1 (Lasso) and L2 (Ridge) penalties. This is particularly effective for omics data where features (proteins) are correlated.

Workflow¶

Data Splitting: The dataset was split into training (e.g., 70%) and testing (e.g., 30%) sets using createDataPartition from the caret package to ensure balanced class distributions.
Feature Scaling: Features were centered and scaled (Z-score) to ensure equal contribution to the model.
Cross-Validation: We performed k-fold cross-validation (e.g., 5-fold) on the training set to optimize the regularization parameter (\(\lambda\)).
Performance Evaluation: The final model was evaluated on the held-out test set using metrics such as AUC (Area Under the ROC Curve), Sensitivity, and Specificity.

# Example: Elastic Net Training with Caret and Glmnet
cv_fit <- cv.glmnet(
  x = as.matrix(X_train),
  y = y_train,
  family = "binomial",
  alpha = 0.5, # Elastic Net mixing parameter
  nfolds = 5
)
best_lambda <- cv_fit$lambda.min

17.7 Robustness and Sensitivity Analysis¶

Script: 08_Supplementary_analysis_v3.Rmd

A critical component of this study was validating the findings. We performed rigorous sensitivity analyses to ensure the results were robust and reproducible.

Module Preservation¶

We tested whether modules discovered in the ADNI cohort were preserved in the independent ADRC cohort using the modulePreservation function from WGCNA. This generates a Z-summary statistic; a Z-summary > 10 indicates strong evidence of preservation.

Adjusted Regression Models¶

To isolate specific signals (e.g., vascular contributions) from general AD pathology, we ran adjusted regression models controlling for covariates such as Age, Sex, and APOE genotype.

# Example: Running Module Preservation (ADNI as Reference)
multiExpr_Recip <- list(
  ADNI = list(data = expr_ADNI),
  ADRC = list(data = expr_ADRC)
)

pres <- modulePreservation(
  multiExpr_Recip,
  multiColor = multiColor_Recip,
  referenceNetworks = 1,
  nPermutations = 500,
  networkType = "signed",
  randomSeed = 123
)

Summary¶

This pipeline transforms raw proteomic data into biologically meaningful modules. By following these four steps—Preprocessing, Network Construction, Module Detection, and Trait Correlation—we identify systems-level protein signatures associated with Alzheimer's Disease.

For clinical and translational research, reproducibility is paramount. Best practices:

R environment management: Use renv to pin package versions for complete reproducibility.
Session documentation: Export sessionInfo() and parameter files at the end of each analysis.
Data deposition: When possible, deposit processed expression matrices (e.g., GEO) and network results with methods metadata.
Validation strategy: Where feasible, repeat key modules in an independent cohort and report module preservation Z-scores.

Publishing reproducible R Markdown notebooks (or targets pipelines) ensures other researchers can verify and extend your findings.