TPOT

TPOT automates optimization of machine learning pipelines using genetic programming to discover and tune predictive models for biomedical and genomic datasets.

Key Features:

• Genetic Programming: Represents machine learning pipelines as expression trees and uses genetic programming to stochastically search and evolve pipeline structures and parameters.
• Pipeline Optimization: Automates selection and tuning of preprocessing steps, feature selection techniques, and classifiers for specific datasets.
• Feature Set Selector (FSS): Allows specification of subsets of features as separate datasets so the genetic programming process can focus on relevant feature sets for large-scale data.
• Template (strongly typed genetic programming): Enforces type constraints and integrates FSS into each pipeline to constrain search space and improve interpretability.
• Performance Metrics: Optimizes pipelines using metrics including accuracy, precision, recall, balanced accuracy, area under the curve (AUC), and area under the precision-recall curve.

Scientific Applications:

• Coronary Artery Disease (CAD) Diagnostics: Applied to predict angiographic diagnoses of CAD from nuclear magnetic resonance-derived lipoprotein and metabolite profiles, outperforming grid search-optimized models on balanced accuracy and area under the precision-recall curve when distinguishing non-obstructive CAD from no CAD.
• Hypertrophic Cardiomyopathy (HCM) Analysis: Used to predict late gadolinium enhancement (LGE) status from native T1 mapping data, achieving high diagnostic accuracy without contrast-enhanced sequences for detection of myocardial fibrosis.
• Genomic Data Analysis: Scaled to RNA-Seq data in a study of major depressive disorder, corroborating associations with depression severity and providing predictive insights into clinical diagnoses.

Methodology:

TPOT uses genetic programming to iteratively evolve machine learning pipelines and optimizes them against performance metrics such as accuracy, precision, recall, balanced accuracy, AUC, and precision-recall AUC, while performing feature selection and handling diverse data types.