EPS

EPS performs feature selection and ranking in high-dimensional biological case-control datasets by combining variational autoencoder-derived latent representations, logistic regression classification, generation of extreme pseudo-samples, and regression modeling on an upsampled space.

Key Features:

• Variational Autoencoder (VAE): Learns complex latent representations from high-dimensional biological input data.
• Logistic Regression Classification: Classifies VAE latent representations into cases and controls to identify discriminative directions in the latent space.
• Extreme Pseudo-sample Generation: Identifies extreme cases and controls from logistic regression outputs and generates pseudo-samples around these extremes to augment the dataset.
• Upsampled Regression Training: Trains a regression model on the combined original and pseudo-samples in the upsampled space to refine variable importance estimates.
• Feature Selection and Ranking: Selects and ranks the most significant variables from the regression model for downstream predictive and interpretative analyses.
• Target Data Types: Designed for high-dimensional biological case-control datasets such as genomics and proteomics.

Scientific Applications:

• Genomics: Identification and prioritization of predictive genomic features and variants in case-control studies.
• Proteomics: Discovery and ranking of informative protein-level features in high-dimensional proteomics datasets.
• Disease Association Studies: Identification of variables associated with disease status in case-control designs.
• Biomarker Discovery: Prioritization of candidate biomarkers via ranked feature importance derived from regression on upsampled data.

Methodology:

Generate latent representations using a variational autoencoder (VAE); classify latent representations with logistic regression to identify extremes; generate pseudo-samples around extreme cases and controls; train a regression model on the upsampled space composed of original and pseudo-samples; select the most significant variables from the regression model.