Detecting SNP markers discriminating horse breeds by deep learning

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate assignment of individuals to their population of origin is crucial for genomic applications.
• High-throughput genotyping generates vast amounts of data, necessitating efficient methods for marker selection.

Purpose of the Study:

• To evaluate Artificial Neural Networks (ANNs), including Deep Neural Networks (DNN), Garson, and Olden methods, for selecting informative SNP markers.
• To assess the performance of these methods in assigning individuals to their true breed of origin.

Main Methods:

• Utilized genomic data from 795 animals across 37 breeds genotyped on an Illumina SNP 50k Bead chip.
• Applied Deep Neural Networks (DNN), Garson, and Olden methods for SNP feature selection.
• Validated marker performance using Principal Component Analysis (PCA), log-likelihood ratios (LLR), and Neighbor-Joining (NJ) trees.

Main Results:

• DNN, Garson, and Olden methods identified reduced SNP panels (4270, 4937, 7999 markers) for breed assignment.
• Achieved 70% assignment success with 110 (DNN), 208 (Garson), and 178 (Olden) markers.
• DNN demonstrated superior performance with 93% accuracy at the highest stringency threshold.
• Selected markers were validated on an independent dataset, correctly assigning all unknown samples.

Conclusions:

• Deep Neural Networks (DNN) offer an efficient strategy for selecting a minimal set of highly discriminant SNP markers.
• The identified SNP panels are effective for accurate breed assignment and tracing the population of origin.
• DNN shows significant potential for feature selection in genomic data analysis.