MinLinMo: a minimalist approach to variable selection and linear model prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Statistical Genetics

Background:

• High-dimensional data analysis often yields complex prediction models.
• Causal inference in large models presents practical challenges.
• Need for efficient and interpretable prediction models in biological research.

Purpose of the Study:

• Introduce MinLinMo, a software package for generating small linear prediction models.
• Prioritize model parsimony, speed, and minimal memory usage.
• Facilitate causal inference with high-dimensional biological datasets.

Main Methods:

• Developed the stand-alone software package MinLinMo.
• Focused on selecting predictors correlated with the outcome.
• Emphasized parsimony, minimal memory footprint, and computational speed.

Main Results:

• MinLinMo successfully generated parsimonious prediction models for epigenetic datasets.
• Models predicted chronological age (15 predictors), gestational age (14 predictors), and birth weight (10 predictors).
• MinLinMo models performed comparably to established models using hundreds of predictors.

Conclusions:

• MinLinMo offers an effective approach to building small, interpretable prediction models.
• Parsimonious models derived from MinLinMo facilitate causal inference in high-dimensional data.
• The software provides a computationally efficient alternative for biological data analysis.