Differentiable Learning of Sequence-Specific Minimizer Schemes with DeepMinimizer.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Minimizers are crucial for sampling k-mers in biological sequence analysis, impacting read mapping and taxonomy prediction.
• Current minimizer schemes often lack optimal density, increasing computational and memory demands.
• Existing methods struggle to optimize minimizer schemes for specific sequences or achieve significant density improvements.

Purpose of the Study:

• To develop a novel method for constructing highly efficient minimizer schemes with optimal density.
• To address the discrete optimization challenge in learning effective minimizer strategies.
• To improve the performance of k-mer sampling for biological sequence analysis.

Main Methods:

• Introduced DeepMinimizer, the first continuous relaxation of the density minimizing objective.
• Employed a novel Deep Learning twin architecture for simultaneous validity and performance.
• Utilized a fully differentiable surrogate objective for efficient gradient-based optimization on GPUs.

Main Results:

• DeepMinimizer successfully discovers minimizer schemes with significantly improved density.
• The method outperforms existing state-of-the-art constructions on human genomic sequences.
• Demonstrated the effectiveness of continuous relaxation and deep learning for this optimization problem.

Conclusions:

• DeepMinimizer offers a powerful and efficient solution for generating optimal minimizer schemes.
• The approach advances k-mer sampling techniques in bioinformatics.
• This work paves the way for more computationally efficient biological sequence analysis tools.