snpnet

snpnet implements scalable penalized regression methods for high-dimensional SNP-based genetic association modeling and polygenic prediction in large cohort genomic datasets.

Key Features:

• Batch Screening (BASIL): Implements the Batch Screening Iterative Lasso (BASIL) framework to enable integration with existing lasso solvers for scalable variable screening on datasets that exceed memory constraints.
• Model Support: Supports ℓ1-penalized linear models, logistic regression, Cox proportional hazards models, and elastic net regularization with ℓ1/ℓ2 penalties.
• Two-bit Genotype Encoding: Uses a two-bit encoding for genetic variants with values {0, 1, 2, NA} to reduce memory relative to double precision representations by a factor of 32.
• Sparse Matrix Encoding: Employs sparse matrix encoding for datasets with many rare variants and uses a simplified compressed sparse block format to exploit sparsity.
• Parallel Computation: Facilitates parallelized matrix–vector multiplications across multiple CPU cores to accelerate computation on large genotype matrices.
• Specialized Solvers: Provides snpnet-2.0, which uses an iteratively reweighted least squares algorithm for Lasso on genetic matrices, and sparse-snpnet, which solves group Lasso problems with an accelerated proximal gradient method.
• Performance Benchmark: Demonstrates the ability to solve regression problems involving up to ~1 million variants and ~100,000 individuals within reported computational bounds (≈10 minutes and <32 GB memory).

Scientific Applications:

• Large-scale cohort GWAS: Enables high-dimensional association analyses in cohorts such as the UK Biobank.
• Genotype–phenotype mapping: Facilitates investigation of associations between SNPs and a wide array of traits and disease outcomes.
• Phenotype prediction and polygenic risk scores: Improves predictive modeling for phenotypes (e.g., height, body mass index, asthma, high cholesterol) and construction of polygenic risk scores.

Methodology:

Computational methods explicitly include the Batch Screening Iterative Lasso (BASIL) framework, integration with existing lasso solvers, ℓ1 and elastic net (ℓ1/ℓ2) penalties, logistic and Cox models, two-bit genotype encoding ({0,1,2,NA}), sparse matrix encoding with a compressed sparse block format, parallelized matrix–vector multiplications, iteratively reweighted least squares for snpnet-2.0, and an accelerated proximal gradient method for group Lasso in sparse-snpnet.