DeepLINK: Deep learning inference using knockoffs with applications to genomics.

Area of Science:

• Computational statistics
• Machine learning
• Genomics

Background:

• High-dimensional data analysis presents challenges in feature selection.
• Controlling the false discovery rate (FDR) is crucial for reliable inference.
• Existing methods may struggle with complex covariate distributions.

Purpose of the Study:

• To introduce DeepLINK, a deep learning framework for FDR-controlled feature selection.
• To ensure applicability across diverse covariate distributions, including nonlinear latent factor models.
• To demonstrate high performance in both simulated and real-world datasets.

Main Methods:

• Developed a two-part deep learning architecture: an autoencoder for knockoff variable construction and a multilayer perceptron for feature selection.
• Implemented FDR control mechanisms within the feature selection process.
• Utilized extensive simulation studies to evaluate performance.

Main Results:

• DeepLINK effectively controls the false discovery rate (FDR) in high-dimensional settings.
• The framework achieves high selection power and prediction accuracy.
• Demonstrated practical utility through successful application to three real-world datasets.

Conclusions:

• DeepLINK offers a robust and effective deep learning solution for FDR-controlled feature selection.
• The method is versatile, handling complex data structures.
• DeepLINK shows significant promise for applications in various scientific domains requiring reliable feature selection.