Learning sparse log-ratios for high-throughput sequencing data.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Discovering sparse biomarkers associated with outcomes is crucial in bioinformatics.
• Log-ratio biomarkers are important for compositional data (CoDa), including high-throughput sequencing (HTS) data.
• Identifying these biomarkers is computationally challenging due to combinatorial optimization problems, limiting current methods' scalability.

Purpose of the Study:

• To present CoDaCoRe, a novel deep learning algorithm for identifying sparse, interpretable, and predictive log-ratio biomarkers.
• To address the computational challenges and scalability limitations of existing methods for biomarker discovery in HTS and CoDa.

Main Methods:

• CoDaCoRe employs a continuous relaxation technique to approximate the combinatorial optimization problem.
• The relaxation is efficiently optimized using gradient descent and the modern machine learning (ML) toolbox.
• The algorithm is designed to handle high-dimensional datasets.

Main Results:

• CoDaCoRe runs orders of magnitude faster than existing methods.
• The algorithm achieves state-of-the-art performance in predictive accuracy and biomarker sparsity.
• Outperformance is validated across diverse datasets, including challenging high-dimensional microbiome, metabolite, and microRNA data.

Conclusions:

• CoDaCoRe offers a computationally efficient and highly accurate solution for log-ratio biomarker discovery.
• The method overcomes scalability limitations, enabling analysis of previously intractable high-dimensional datasets.
• CoDaCoRe represents a significant advancement in the automated discovery of sparse biomarkers from complex biological data.