HARVESTMAN: a framework for hierarchical feature learning and selection from whole genome sequencing data

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Supervised learning with high-throughput sequencing data faces challenges like the curse of dimensionality, leading to overfitting and scalability issues.
• Genomic variant calls may not be optimal for specific learning tasks, hindering predictive accuracy.
• Existing methods struggle with efficient feature selection from complex genomic datasets.

Purpose of the Study:

• To introduce HARVESTMAN, a method for automatic feature learning, selection, and model building using hierarchical relationships of genomic variants.
• To address the limitations of current approaches in handling large-scale genomic data and variant representations.
• To improve the accuracy and efficiency of predictive models built from high-throughput sequencing data.

Main Methods:

• HARVESTMAN leverages hierarchical relationships among biological interpretations of genomic variants.
• It constructs a knowledge graph over genomic variants.
• An integer linear program is solved to find optimal genomic variant encodings.

Main Results:

• HARVESTMAN scales to thousands of genomes with over 84 million variants, demonstrated on 1000 Genomes Project data.
• It outperforms binary SNP representation for breast cancer data from The Cancer Genome Atlas.
• HARVESTMAN selects more parsimonious and less redundant feature subsets compared to existing methods while maintaining classifier accuracy.

Conclusions:

• HARVESTMAN is an effective hierarchical feature selection approach for supervised learning from variant call data.
• The method automatically identifies optimal genomic variant encodings through knowledge graphs and integer linear programming.
• HARVESTMAN is faster and more parsimonious than other hierarchical feature selection methods.