A MINE alternative to D-optimal designs for the linear model.

Area of Science:

• Genomics
• Statistical Modeling
• Experimental Design

Background:

• Large-scale genomics experiments are costly, necessitating efficient information extraction.
• The Maximally Informative Next Experiment (MINE) criterion guides experimental design to maximize information gain.
• Exploring MINE in the context of linear models offers a simplified yet relevant framework.

Purpose of the Study:

• To adapt and evaluate four variations of the MINE criterion for linear models.
• To theoretically establish conditions for MINE criterion maximization and its equivalence to D-optimality.
• To assess the performance of MINE variations in identifying true linear relationships and controlling false positives.

Main Methods:

• Development of four MINE variations: MINE-like, MINE, MINE with random orthonormal basis, and MINE with random rotation.
• Theoretical analysis using Theorem 1 for maximization conditions and Theorem 2 for D-optimality equivalence.
• Simulation studies under linear models with varying dimensions (p >> n) and sample sizes (n < 100).

Main Results:

• Two MINE variations (random orthonormal basis, random rotation) demonstrated faster discovery of true linear relations in high-dimensional settings (p >> n).
• These MINE variations exhibited lower false positive rates compared to MINE-like and, often, MINE methods in simulations.
• Sufficient conditions for MINE maximization and its link to D-optimality were theoretically established.

Conclusions:

• MINE variations, particularly those employing random bases or rotations, offer enhanced efficiency and accuracy in linear model-based experimental design.
• These findings are crucial for optimizing resource allocation in large-scale genomics and other high-dimensional studies.
• The theoretical and simulation results provide a strong foundation for applying MINE in practical experimental design.