Transfer learning for genotype-phenotype prediction using deep learning models.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genotype data is often scarce for understudied populations, hindering genotype-phenotype prediction.
• Knowledge from large populations can be transferred to improve predictions in smaller, data-sparse groups.

Purpose of the Study:

• To illustrate the applicability of transfer learning for genotype data and genotype-phenotype prediction.
• To develop a pipeline for transferring knowledge for case/control classification in small populations.

Main Methods:

• Generated eight phenotypes using HAPGEN2 and PhenotypeSimulator for large (CEU) and small (YRI) populations.
• Evaluated the method with 5 and 10 risk single nucleotide polymorphisms (SNPs) for different phenotypes.

Main Results:

• Transfer learning improved classification accuracy by 2-14.2% across eight phenotypes.
• Statistically significant improvements were observed (p=0.0306 for 5 SNPs, p=0.0478 for 10 SNPs).

Conclusions:

• The proposed pipeline effectively transfers knowledge for genotype-phenotype prediction in small populations.
• Transfer learning shows promise for personalized medicine and endangered species genomics.
• Larger source population data significantly boosts transfer learning performance.