Log-ratio lasso: Scalable, sparse estimation for log-ratio models.

Area of Science:

• Biological and medical sciences
• Biostatistics
• Proteomics

Background:

• Positive-valued signal data is prevalent in biological and medical sciences, often generated by techniques like mass spectrometry.
• The meaningful information in this data lies in the relative intensities, making log-ratios of features a key analytical approach.
• Analyzing all pairwise log-ratios leads to high-dimensional predictor spaces, necessitating computationally efficient methods.

Purpose of the Study:

• To develop an efficient computational procedure for analyzing high-dimensional log-ratio data.
• To introduce a method for fitting sparse models that are both interpretable and predictive.
• To improve upon existing methods for analyzing positive-valued biological and medical data.

Main Methods:

• An embedding of the log-ratio parameter space into a lower-dimensional space was developed.
• An efficient penalized fitting procedure was created using this embedding.
• A two-step fitting procedure combining convex filtering and non-convex pruning was implemented to achieve sparsity.

Main Results:

• The proposed method successfully fits highly sparse models to cancer proteomics data.
• The identified features in the sparse model possess known biological relevance.
• The method significantly enhances predictive accuracy compared to less interpretable approaches.

Conclusions:

• The developed embedding and fitting procedure provide an efficient way to handle high-dimensional log-ratio data in biological and medical research.
• This approach yields sparse, interpretable, and highly predictive models.
• The method demonstrates strong performance on a cancer proteomics dataset, highlighting its utility in biomarker discovery and predictive modeling.