Multikernel linear mixed models for complex phenotype prediction.

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Linear mixed models (LMMs) are standard for complex trait phenotype prediction.
• Current LMMs are limited by assumptions of simple genetic architectures.
• Modeling complex genetic interactions remains a challenge.

Purpose of the Study:

• To introduce a novel predictive modeling framework, multikernel linear mixed model (MKLMM), extending LMMs.
• To enable the modeling of genetic interactions, especially local interactions between nearby variants.
• To develop MKLMM-Adapt for automatic inference of interaction types across genomic regions.

Main Methods:

• Developed MKLMM, integrating multiple-kernel machine learning with LMMs.
• Implemented MKLMM-Adapt for automated genomic interaction type inference.
• Applied MKLMM and MKLMM-Adapt to analyze case-control and mouse phenotype datasets.

Main Results:

• MKLMM-Adapt demonstrated superior performance in phenotype prediction compared to competing methods.
• The framework effectively models complex genetic interactions.
• MKLMM showed comparable computational efficiency to standard LMMs.

Conclusions:

• MKLMM and MKLMM-Adapt offer a powerful and computationally feasible approach for phenotype prediction.
• The methods advance the ability to model complex genetic architectures and interactions.
• This framework achieves state-of-the-art predictive power without compromising computational feasibility or genomic privacy.