Harvestman

Harvestman performs hierarchical feature learning and selection from whole genome sequencing (WGS) data to generate biologically informed feature representations for genomic classification and interpretation.

Key Features:

• Hierarchical Feature Learning: Exploits hierarchical relationships among genomic variant representations to perform automatic feature learning that captures complex biological structure.
• Scalability: Processes large-scale WGS datasets, demonstrated on 1000 Genomes Project phase 3 data containing over 84 million variants and thousands of genomes.
• Adaptive Feature Selection: Selects and combines diverse feature representations tailored to specific learning tasks, as shown in analyses of TCGA breast cancer data.
• Efficient Feature Subset Selection: Identifies smaller, less redundant feature subsets while maintaining or improving classifier accuracy relative to traditional binary SNP representation methods.

Scientific Applications:

• Genomic Variant Analysis: Enables population-scale variant analysis using datasets such as the 1000 Genomes Project phase 3 to study human genetic diversity and its implications for health.
• Cancer Research: Applied to The Cancer Genome Atlas (TCGA) breast cancer data to identify genomic features associated with cancer phenotypes.

Methodology:

Harvestman integrates hierarchical relationships among genomic variants into the feature learning and selection process, producing nuanced representations that facilitate improved model performance and interpretability.