Empowering individual trait prediction using interactions for precision medicine.

Area of Science:

• Genetic Epidemiology
• Computational Biology
• Biostatistics

Background:

• Precision medicine aims to enhance individual risk prediction using complex models.
• Polygenic diseases often involve interactions between biological and genetic features.
• Standard statistical models may be suboptimal for omics data, necessitating machine learning approaches.

Purpose of the Study:

• To develop an advanced model-based multifactor dimensionality reduction (MB-MDR) algorithm for improved individual prediction.
• To enable prediction models that explicitly incorporate feature interactions.

Main Methods:

• Extension of the original multifactor dimensionality reduction (MDR) algorithm.
• Development of a model-based MDR (MB-MDR) approach for interaction-enhanced prediction.
• Utilized a comprehensive simulation study and real-world rheumatoid arthritis dataset.

Main Results:

• The new MB-MDR algorithm achieved a median AUC of 0.66, outperforming Random Forest (0.54) and Elastic Net (0.50) when feature interactions were present.
• Algorithm performance was comparable to other methods when no interactions were detected.
• Demonstrated applicability to real-world data, including a rheumatoid arthritis dataset.

Conclusions:

• Explicitly modeling feature interactions significantly enhances prediction performance.
• Incorporating interactions is crucial for advancing precision medicine.
• The developed MB-MDR algorithm is applicable to various datasets with discrete features and is available as an R package.