Genome-wide association data classification and SNPs selection using two-stage quality-based Random Forests.

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Genome-wide association data analysis involves selecting single-nucleotide polymorphisms (SNPs), a challenging task due to high dimensionality and irrelevant data.
• Random Forests (RF) are effective for identifying genetic variants in complex diseases but struggle with SNP selection and model accuracy in Genome-wide Association Studies (GWAS).

Purpose of the Study:

• To propose a new two-stage quality-based sampling method in Random Forests (ts-RF) for improved SNP subspace selection in GWAS.
• To enhance the accuracy of prediction models and reduce the dimensionality of GWAS data.

Main Methods:

• A two-stage quality-based sampling method (ts-RF) was developed for SNP subspace selection in GWAS.
• The method uses p-value assessment to categorize SNPs into informative and irrelevant groups, further dividing informative SNPs into highly and weakly informative subgroups for sampling.

Main Results:

• The ts-RF approach generated more accurate trees with lower prediction error, potentially avoiding overfitting.
• Experiments on Parkinson's and Alzheimer's datasets demonstrated significant reduction in prediction errors, outperforming existing state-of-the-art random forests.
• The model identified top SNPs, including genes associated with neurological disorders, and reduced data dimensionality.

Conclusions:

• The ts-RF method effectively selects informative SNP subgroups associated with diseases, outperforming traditional statistical approaches.
• This RF model is well-suited for GWAS data where the number of cases is significantly smaller than the number of SNPs.
• Experimental results confirm the superiority of ts-RF over existing RF methods like Breiman's RF, GRRF, and wsRF.