PUMA: a unified framework for penalized multiple regression analysis of GWAS data

Area of Science:

• Genetics
• Statistical genetics
• Computational biology

Background:

• Penalized Multiple Regression (PMR) methods aim to discover novel disease associations in genome-wide association studies (GWAS).
• Existing PMR methods face challenges with computational speed, performance on large datasets, and biological plausibility of identified associations, limiting their application.

Purpose of the Study:

• To develop and validate the Penalized Unified Multiple-locus Association (PUMA) analysis framework.
• To address limitations of previous PMR methods, including computational efficiency and robust identification of disease-associated loci.

Main Methods:

• Developed a combined algorithmic and heuristic framework for PUMA analysis.
• Implemented a novel minorize-maximization (MM) algorithm for generalized linear models (GLM).
• Integrated heuristic model selection and testing with various penalized maximum likelihood penalties (Lasso, Adaptive Lasso, NEG, MCP, LOG).

Main Results:

• PUMA demonstrated high performance in simulations mirroring real GWAS data, reliably increasing power to detect weak associations.
• PUMA outperformed existing PMR methods, which sometimes performed worse than single marker testing.
• Analysis of GWAS data for type 1 diabetes, Crohn's disease, and rheumatoid arthritis replicated known associations and identified novel susceptibility loci.

Conclusions:

• The PUMA framework offers a computationally efficient and powerful approach for identifying novel genetic associations in GWAS.
• PUMA successfully identified novel, etiologically relevant loci for type 1 diabetes, Crohn's disease, and rheumatoid arthritis, which were undetectable by standard methods.
• The study provides software for applying the PUMA analysis framework, facilitating broader application in genetic research.