Metric learning for comparing genomic data with triplet network.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Biological applications often require pairwise comparisons of complex data.
• Existing dissimilarity metrics may not be universally applicable across diverse biological datasets.
• Developing adaptive metrics is crucial for accurate data analysis.

Purpose of the Study:

• To introduce MEtric Learning with Triplet network (MELT), a novel framework for learning data-adaptive dissimilarity metrics.
• To enable accurate pairwise comparisons in biological applications where labeled data is scarce.
• To provide a flexible and robust solution for large-scale genomic data analysis.

Main Methods:

• Developed MELT, a weakly supervised, data-driven metric learning framework using a Triplet network.
• MELT learns a nonlinear mapping to an embedding space, optimizing similarity and dissimilarity.
• Applied MELT to genomic sequence comparison, microbiome analysis, and single-cell gene expression profiling.

Main Results:

• MELT demonstrated superior empirical utility compared to widely used dissimilarity metrics in experiments.
• The framework successfully learned adaptive dissimilarity for datasets lacking explicit grouping information.
• Achieved more accurate and adaptive dissimilarity measures in complex biological comparisons.

Conclusions:

• MELT offers a powerful approach for learning adaptive dissimilarity metrics in biological pairwise comparisons.
• The framework is particularly valuable for applications lacking labeled data.
• MELT is expected to enhance a broad range of large-scale genomic comparison tasks.