Identifying genetic determinants of complex phenotypes from whole genome sequence data

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Machine Learning Applications

Background:

• Relating genotype to phenotype is a key biological challenge, especially for complex traits.
• Traditional methods like GWAS using Single Nucleotide Polymorphism (SNP) data have limitations.
• The efficacy of machine learning (ML) on whole proteome data for identifying genetic determinants is underexplored.

Purpose of the Study:

• To evaluate the applicability of ML algorithms for identifying genetic determinants in whole proteome data.
• To develop and assess a 'chunking' layer to enhance the analysis of large proteome datasets.
• To compare the performance of Adaptive Boosting (AB) and Repeated Random Forest (RRF) algorithms.

Main Methods:

• Implementation and tuning of Adaptive Boosting (AB) and Repeated Random Forest (RRF) ML algorithms.
• Development of a chunking layer to facilitate whole proteome data analysis.
• Performance assessment using influenza data (known determinants) and Pseudomonas aeruginosa data (drug resistance).

Main Results:

• Chunking improved runtime by an order of magnitude and increased prediction sensitivity.
• RRF demonstrated higher sensitivity than AB, especially at smaller chunk sizes.
• RRF provided more accurate predictions with lower error rates for bacterial drug resistance compared to AB.

Conclusions:

• ML algorithms are effective for identifying genetic determinants in small proteomes, such as viruses.
• The developed RRF algorithm shows promise for complex trait analysis in genomics.
• Decreasing costs of sequencing and phenotyping support further investigation of these ML approaches.