Amalgams: data-driven amalgamation for the dimensionality reduction of compositional data.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Analysis

Background:

• Next-generation sequencing data often contains only relative transcript information due to biological and technical limitations.
• Interpreting individual components in relative data is challenging, necessitating specialized analytical approaches.
• Compositional data analysis (CoDa) offers methods using log-ratio transforms for relative data.

Purpose of the Study:

• To develop a novel dimension reduction technique for compositional data that is both powerful and interpretable.
• To address the challenge of high dimensionality in compositional datasets where features often outnumber samples.
• To create a method that leverages the non-linear distortions inherent in amalgamation for improved data analysis.

Main Methods:

• Proposing data-driven amalgamation, a method that finds optimal amalgamations to reduce dimensionality.
• Implementing the method in the R package 'amalgam' for user-friendly application.
• Benchmarking the method on 13 real-world datasets against state-of-the-art techniques.

Main Results:

• Data-driven amalgamation effectively reduces dimensionality in compositional data.
• The method optimally preserves distances between samples or aids in classifying samples as diseased or not.
• Performance on benchmark datasets is competitive with existing state-of-the-art methods.

Conclusions:

• Data-driven amalgamation provides a powerful and interpretable approach to dimensionality reduction for compositional data.
• The method generates new features that are easily understood as sums of original parts.
• The 'amalgam' R package offers a practical tool for researchers working with high-dimensional relative sequencing data.