wenda

wenda performs unsupervised domain adaptation by applying a weighted elastic net to prioritize features that are invariant across domains and improve predictive accuracy for high-dimensional biological datasets with distribution mismatches.

Key Features:

• Unsupervised Domain Adaptation: Operates without labeled data in the target domain to adapt models across distribution mismatches.
• Weighted Elastic Net Regularization: Incorporates a weighted elastic net penalty that adjusts feature penalties based on cross-domain behavior to prioritize consistent features.
• Handling High-Dimensional Data: Suited for p >> n settings common in biological data such as DNA methylation profiles, enabling feature selection and regularization under high dimensionality.

Scientific Applications:

• Age Prediction from DNA Methylation: Predicts chronological age from epigenetic DNA methylation profiles across multiple tissues while adapting to tissue-specific distribution shifts.
• Cross-Tissue Prediction and Cerebellum Robustness: Reduces prediction errors for tissues not represented in training data, notably improving performance on cerebellum samples.
• Evaluation on Simulated and Real Data: Applied to simulated datasets with controlled distribution mismatches and real-world DNA methylation datasets to quantify error reduction.

Methodology:

The method compares dependencies between input features in training and test datasets and implements a weighted elastic net that penalizes features showing different behavior across domains.