Seagull: lasso, group lasso and sparse-group lasso regularization for linear regression models via proximal gradient...

Area of Science:

• Life Sciences
• Bioinformatics
• Computational Biology

Background:

• High-dimensional linear models are common in biological data analysis.
• Penalization methods like lasso are crucial for handling multicollinearity.
• Existing methods can be computationally intensive.

Purpose of the Study:

• To introduce a fast and numerically efficient R package, seagull, for penalized linear regression.
• To implement lasso, group lasso, and sparse-group lasso operators.
• To enable the computation of complete regularization paths.

Main Methods:

• Proximal gradient descent for efficient computation of penalized operators.
• Warm starts for fast grid search of penalty parameters.
• Backtracking line search for determining step size in iterations.

Main Results:

• The seagull package demonstrates comparable accuracy to the SGL package in predicting biological age from methylation data (R² > 0.99).
• seagull computes solutions significantly faster than SGL.
• The package supports feature weighting, enhancing its flexibility.

Conclusions:

• seagull provides a comprehensive suite of lasso variants, including lasso, group lasso, sparse-group lasso, and IPF-lasso.
• It offers a computationally efficient and flexible tool for high-dimensional linear modeling in biological sciences.
• The package facilitates precise biological age prediction from DNA methylation data.