Functional embedding for the classification of gene expression profiles.

Area of Science:

• Genomics
• Statistical Genetics
• Data Science

Background:

• High-dimensional data (large p, small n) presents challenges for traditional multivariate analysis due to ill-conditioned variance-covariance matrices.
• Functional data analysis (FDA) offers potential for analyzing high-dimensional data due to its design for infinite-dimensional data.
• Existing methods struggle with the complexities of genomics data characterized by a high number of features and limited samples.

Purpose of the Study:

• Propose a novel functional embedding (FEM) technique to bridge multivariate and functional data analysis.
• Address the challenges posed by high dimensionality (p) in genomic and statistical genetics data.
• Enhance analytical capabilities by borrowing strength across samples using FDA techniques.

Main Methods:

• Develop a functional embedding (FEM) technique using pairwise dissimilarities among predictor variables.
• Create a univariate configuration of covariates, interpreted as variable ordination.
• Define the domain of a function space for FEM, enabling downstream functional logistic regression for classification.

Main Results:

• Successfully applied FEM to classify high-dimensional multivariate data, demonstrated via functional logistic regression.
• Evaluated FEM on gene expression array and mass spectrometry datasets.
• Achieved favorable comparison against previously employed methods for high-dimensional gene expression profile classification.

Conclusions:

• The proposed functional embedding (FEM) technique effectively handles high-dimensional data in genomics.
• FEM provides a robust approach for classification tasks in statistical genetics.
• This method offers a promising alternative for analyzing complex biological datasets with n<