Multi-Modality Machine Learning Predicting Parkinson’s Disease (Online Appendix)

LNG ❤️ Open Science 😍

Summary

This is the online appendix for the manuscript titled "Multi-Modality Machine Learning Predicting Parkinson’s Disease", where we integrated clinico-demographic, genetic, and transcriptomic data within an automated machine learning open science framework (GenoML) to predict Parkinson’s disease and identify potential novel therapeutic targets for drug development

Last Updated: June 2021

Workflow Diagram

Helpful Links

• BioRxiv pre-print
• streamlit interactive app
  • streamlit interactive app repository
• GenoML short paper
  • GenoML website
  • GenoML repository

Orientation

• The scripts/ directory includes the pre-processing, munging, training, tuning, optmiziation, and networks scripts (coming soon!)
• The topmodel_G1E5T1E2/ directory includes the
  • Top model's figures, metrics, and .joblib model
  • Genetics at a p-value threshold of 1E-5 only figures, metrics, and .joblib model
  • Transcriptomics at a p-value threshold of 1E-2 only figures, metrics, and .joblib model
  • Clinicodemographic only figures, metrics, and .joblib model
• The shapAnalysis/ directory includes the
  • Surrogate XGBoostClassifier models for the combined model, rsIDs only, ENSGs only, or rsIDs and ENSGs only
  • Figures showing top 5% of features in the combined model
  • A directory with dependence plots between each SNP and its effect on PRS90
• The trained_joblibs/ directory includes all the 49 trained models in .joblib format
• The tuned_joblibs/ directory includes all the 49 tuned models in .joblib format
• The trained_metricsplots/ includes all metrics for the 49 trained models
  • *.trainedModel_withheldSample_probabilities.png
  • *.trainedModel_withheldSample_ROC.png
  • *.training_withheldSamples_performanceMetrics.csv
• The tuned_metricsplots/ directory includes all metrics for the 49 tuned models
  • *.tunedModel_CV_Summary.csv
  • *.tunedModel_top10Iterations_Summary.csv
• The tested_metricsplots/ directory includes all metrics for the 49 tested models in PDBP
  • *.testedModel_allSample_probabilities.png
  • *.testedModel_allSample_ROC.png
  • *.testedModel_allSamples_performanceMetrics.csv
• The tables/ directory includes all the tables referenced in the manuscript and supplemental tables

Sheet Tab Name	Description
T1	Table 1: Descriptive statistics of studies included from AMP-PD.
T2	Table 2: Performance metric summaries comparing training in withheld samples in PPMI
T3	Table 3: Performance metric summaries comparing at tuned cross-validation in withheld samples in PPMI
T4	Table 4: Performance metric summaries comparing combined tuned and untuned model performance on PDBP validation dataset
T5	Table 5: Optimizing the AUC threshold in withheld training samples and in the validation data
ST1	Supp. Table 1: Complete performance metrics for best combined method comparing training in withheld samples in PPMI
ST2	Supp. Table 2: Rarer coding variant burden analyses for genes under GWAS peaks
ST3	Supp. Table 3: Complete summary statistics for QTL Mendelian randomization