AI2AMP-PD

Analysis code for a study of RNA-seq–based differential expression (DE) and machine-learning (ML) classification of Parkinson's disease (PD) case/control status using data from the AMP-PD (Accelerating Medicines Partnership – Parkinson's Disease) cohorts PPMI and PDBP/BioFIND (PDBF).

Project Overview

The pipeline in this repository:

1. Filters and prepares AMP-PD RNA-seq quantification and clinical/genetic data for case/control and mutation-carrier comparisons (DE/src/filtration.py, DE/src/extract_quant.py).
2. Runs differential expression analysis with DESeq2 on gene-level, eRNA, and circRNA quantifications, adjusting for covariates (age, sex, plate, RIN, genotype PCs) (DE/src/DE_PCs.R, DE/src/eRNA/DE_eRNA.R, DE/src/circRNA/DE_circRNA.R).
3. Performs functional enrichment (GO/KEGG/Reactome/GSEA) on DE results (DE/src/Enrich_Profiler.R).
4. Trains and evaluates ML classifiers (logistic regression, LASSO, SVM, random forest, XGBoost, KNN, MLP, etc.) that predict case/control status from DE genes, eRNAs, circRNAs, clinical variables, and polygenic risk scores (PRS), with PPMI used for training and PDBF used as an independent test/replication set (ML/src/*.py).
5. Summarizes model performance with AUROC/PR curves and feature-importance plots (ML/src/31_plots_AUC_PR_inOne.py).
6. Also includes a standalone NanoString endogenous/housekeeping gene QC script (DE/NanoStringData/NanoString-AllControl-Boxplot.R).

Input data: gene/eRNA/circRNA read-count and TPM matrices, sample covariate tables (*_cov_with_PCs.tsv), and AMP-PD clinical CSVs (demographics, UPDRS, MMSE, participant mutation status). Example covariate/feature tables used by the ML scripts are provided under AI2AMP-PD_Datasets/ in the parent directory (PPMI_PDBP/, ML/).

Outputs: DESeq2 result tables (DEresult*.xls), volcano/diagnostic plots, enrichment tables and plots, per-model prediction tables and AUROC/PR plots, and feature-importance/coefficient tables — all written under each module's results/ directory.

Documentation and Working Example

Differential expression (genes)

cd AI2AMP-PD/DE/src
Rscript DE_PCs.R RNA_expr_matrix_tpm.txt RNA_expr_matrix_reads.txt sample_covariates_with_PCs.tsv PDBF_CaseA_CtrlA
# Output: ../results/PDBF_CaseA_CtrlA_PCs/DE/

Differential expression (eRNA / circRNA)

cd AI2AMP-PD/DE/src/eRNA
Rscript DE_eRNA.R          # edit prefix / expr_file_cts at top of script

cd AI2AMP-PD/DE/src/circRNA
Rscript DE_circRNA.R PDBF_all_samples merge.exon.allSamples.PDBF.circRNArawcount.wide.txt

Functional enrichment

cd AI2AMP-PD/DE/src
Rscript Enrich_Profiler.R   # edit the `prefix` variable to match a DE results folder
# Output: ../results/<prefix>/Enrich_profiler/

ML classification

cd AI2AMP-PD/ML/src
python 20_Classifers_DT_SVM.py DT mm          # feature_selection, scale_method
python 11_Classifers_LASSO.py                 # LASSO feature-selection pathway
python 11_Classifers_LASSO_stepped.py
python 11_Classifers_Genes_clinical_PRS_stepadd.py DT logz DEG
python 31_plots_AUC_PR_inOne.py               # combined AUROC/PR summary plots

Each ML script trains on PPMI (data/PPMIvst_DEG_eRNA_circRNA_Clinical_PRS.tsv) and evaluates on held-out PPMI samples plus the independent PDBF cohort (data/PDBFvst_DEG_eRNA_circRNA_Clinical_PRS.tsv), writing per-model train/val/test prediction tables and ROC/PR plots to ../results/.

TIP: The AMP-PD clinical and genetic data used here contain protected health information and are not included in this repository. Access is requested through the AMP-PD Knowledge Portal. Substitute your own similarly-formatted matrices/covariate tables to run the pipeline end to end.

Scripts for Making Figures

Figure/plot generation is embedded in the DE and ML analysis scripts above rather than kept in separate figure-only scripts — running an analysis script produces its associated plots as a side effect, written under that module's results/ directory.

Script	Plots produced	Output
DE/src/DE_PCs.R	Data-diagnosis plots, sample-distance clustering tree, PCA plots, MA plots, volcano plots (EnhancedVolcano), top-DE-gene heatmaps	data_diagnosis.pdf, clustering.tree.pdf, DEresult.padj_05_volcano.pdf, DEresult.padj_05.heatmap.pdf in ../results/<prefix>/DE/
DE/src/eRNA/DE_eRNA.R	Same plot types as DE_PCs.R, for eRNA quantifications	../results/<prefix>/DE/
DE/src/circRNA/DE_circRNA.R	Data-diagnosis, MA, volcano, and top-DE heatmap plots for circRNA quantifications	../results/<prefix>/DE/
DE/src/Enrich_Profiler.R	GO/KEGG/Reactome/GSEA enrichment bar plots	Enrich.pdf in ../results/<prefix>/Enrich_profiler/
DE/NanoStringData/NanoString-AllControl-Boxplot.R	Boxplots of NanoString endogenous/housekeeping control probes (QC/validation)	DE/NanoStringData/results/
ML/src/20_Classifers_DT_SVM.py	Prediction-distribution plots, per-model train/val/test AUROC and PR curves	*_prediction_distribution.pdf, Train_Val_AUROC_PR_<model>.pdf, Test_AUROC_PR_<model>.pdf in ../results/
ML/src/11_Classifers_LASSO.py	LASSO coefficient-vs-alpha and performance-vs-alpha (AUC/accuracy) plots	coefficients_alpha*.pdf, alpha_performances_*.pdf in ../results/
ML/src/11_Classifers_LASSO_summary.py	Best-alpha, feature-count-vs-alpha, and max-correlation-vs-alpha summary plots	best_alpha_loop_LN.pdf, feature_count_alpha_loop_LN_*.pdf, max_corr_alpha_loop_LN.pdf in ../results/
ML/src/11_Classifers_Genes_clinical_PRS_stepadd.py	Same prediction-distribution and AUROC/PR plots as 20_Classifers_DT_SVM.py, for the stepwise gene/clinical/PRS feature-addition models	../results/
ML/src/31_plots_AUC_PR_inOne.py	Combined AUROC and PR curves across all models/cohorts, top-20 feature-importance bar plot	AUROC_merged.pdf, PRC_merged.pdf, Top20_feature_importance_bar_Symbol.pdf in ../results/

Run the corresponding script from its src/ directory (see Documentation and Working Example above) to regenerate its plots; no separate make-figures step is needed.

System Requirements & Dependencies

• OS: Developed and tested on Linux/macOS; R and Python scripts have no OS-specific dependencies.
• R (≥ 4.0): tidyverse, RCurl, hexbin, pheatmap, RColorBrewer, hwriter, vsn, DESeq2, ReportingTools, BiocParallel, limma, EnhancedVolcano, ggplot2, clusterProfiler, biomaRt, KEGG.db, ReactomePA, DOSE, org.Hs.eg.db, genefilter, GO.db, topGO, dplyr, gage, ggsci, enrichplot, fgsea, httr.
• Python (≥ 3.8): pandas, numpy, scikit-learn, matplotlib, seaborn.
• No other versions have been tested.

Recorded computational environment: see session_info/ for the exact R session info (R_sessionInfo.md), Python session info (python_sessionInfo.md), and pinned Python dependencies (requirements.txt) needed to recreate the environment the code was run in. (These currently contain placeholders pending capture of the original run's sessionInfo() / pip freeze output — see that folder for instructions to regenerate them.)

Installation

git clone <this-repository-url>
cd AI2AMP-PD

# R dependencies (auto-installed on first run of DE_PCs.R; for the others, install manually)
Rscript -e 'install.packages(c("tidyverse","RCurl","hexbin","pheatmap","RColorBrewer","hwriter")); \
            if (!requireNamespace("BiocManager", quietly=TRUE)) install.packages("BiocManager"); \
            BiocManager::install(c("vsn","DESeq2","ReportingTools","BiocParallel","limma","EnhancedVolcano", \
                                    "clusterProfiler","biomaRt","KEGG.db","ReactomePA","DOSE","org.Hs.eg.db", \
                                    "GO.db","topGO","gage","enrichplot","fgsea"))'

# Python dependencies
pip install pandas numpy scikit-learn matplotlib seaborn

Folder Structure

AI2AMP-PD/
├── DE/                             # Differential expression analysis
│   ├── src/
│   │   ├── filtration.py           # Build case/control & mutation-carrier sample sets from clinical/genetic data
│   │   ├── extract_quant.py        # Extract per-cohort RNA-seq quantification matrices
│   │   ├── DE_PCs.R                # DESeq2 DE analysis for genes, adjusted for covariates/genotype PCs
│   │   ├── Enrich_Profiler.R       # GO/KEGG/Reactome/GSEA enrichment on DE results
│   │   ├── eRNA/DE_eRNA.R          # DESeq2 DE analysis for eRNAs
│   │   └── circRNA/DE_circRNA.R    # DESeq2 DE analysis for circRNAs
│   ├── NanoStringData/             # NanoString validation data and QC boxplot script
│   └── results/                    # DE tables, volcano/diagnostic plots, enrichment outputs (generated)
├── ML/                             # Machine-learning classification
│   ├── src/
│   │   ├── 11_Classifers_LASSO*.py             # LASSO-based feature selection & classification
│   │   ├── 11_Classifers_Genes_clinical_PRS_stepadd.py  # Stepwise addition of gene/clinical/PRS features
│   │   ├── 20_Classifers_DT_SVM.py             # Multi-model classification (SVM, RF, XGB, KNN, MLP, etc.)
│   │   └── 31_plots_AUC_PR_inOne.py            # Combined AUROC/PR summary plots
│   └── results/                    # Per-model prediction tables, AUROC/PR plots, feature-importance tables (generated)
├── session_info/                   # R/Python session info & dependency versions for reproducing the environment
└── .gitignore

File naming convention: result folders are named <Cohort>_<Comparison>_<covariates> (e.g. PDBF_CaseA_CtrlA_PCs, PPMI_all_samples_circRNAexon); Case/Ctrl refer to PD case vs. control status, and suffixes such as _PCs, _eRNA_Class1, or _circRNAexon indicate the covariate set or feature type used.

Limitations

• Scripts were written for the specific AMP-PD sample sets and covariate schemas used in this study (e.g., case_control_other_latest, genotype PC1–PC10) and are not general-purpose tools; column names in input files must match what each script expects.
• File paths and cohort-specific parameters (prefix, expr_file_cts, etc.) are set as literals or commandArgs/sys.argv inputs at the top of each script and may need to be edited directly for new datasets.
• Raw AMP-PD data are not distributed with this repository; the pipeline cannot be run end-to-end without independently obtained access.

Citation

This code accompanies the manuscript:

Hu R, Dong X, et al., "Differential expression and machine-learning classification of Parkinson's disease using AMP-PD RNA-seq data" (see revision_R1/ for the manuscript and reporting materials).

Principal Investigator: Xianjun Dong, PhD (xianjun.dong@yale.edu)

Please cite the associated publication when using this code.

License

This project is licensed under the MIT License — see LICENSE.txt for details.

Acknowledgements

Data were derived from the AMP PD Knowledge Platform. Funded by Aligning Science Across Parkinson's (ASAP) and the Michael J. Fox Foundation for Parkinson's Research (MJFF).