Molecular group and correlation guided structural learning for multi-phenotype prediction.

Area of Science:

• Bioinformatics
• Genomics
• Statistical Genetics

Background:

• Genomic data analysis often involves complex, multi-phenotypic responses.
• Existing methods may not fully leverage biological grouping or correlation structures.
• Accurate feature selection and outcome prediction are crucial for omics studies.

Purpose of the Study:

• To develop a novel supervised learning tool, Brilliant, for feature selection and outcome prediction in high-dimensional genomic data.
• To incorporate genome and phenotype grouping and correlation structures into a penalized multi-response linear regression model.
• To evaluate Brilliant's performance against existing methods and apply it to real-world omics datasets.

Main Methods:

• Biological gRoup guIded muLtivariate muLtiple lIneAr regression with peNalizaTion (Brilliant) model.
• Penalized multi-response linear regression incorporating grouping and correlation structures.
• Extensive simulations and application to two omics studies (asthma cohort, cell-type deconvolution).

Main Results:

• Brilliant demonstrated superior performance in simulations compared to competing methods.
• Identified novel associations between gene expression and DNA methylation in an asthma cohort.
• Improved cell-type fraction prediction accuracy and discovered novel signature genes in a gene expression study.

Conclusions:

• Brilliant is an effective bioinformatics tool for feature selection and outcome prediction in multi-phenotypic genomic data.
• The tool provides valuable biological insights by integrating structural information.
• Brilliant advances the analysis of complex omics data, with applications in disease research and cell-type deconvolution.