Machine learning as an effective method for identifying true SNPs in polyploid plants
Single Nucleotide Polymorphisms (SNPs) have many advantages as molecular markers since they are ubiquitous and co-dominant. However, the discovery of true SNPs especially in polyploid species is difficult. Peanut is an allopolyploid, which has a very low rate of true SNP calling. A large set of true and false SNPs identified from the Arachis 58k Affymetrix array was leveraged to train machine learning models to select true SNPs straight from sequence data. These models achieved accuracy rates of above 80% using real peanut RNA-seq and whole genome shotgun (WGS) re-sequencing data, which is higher than previously reported for polyploids. A 48K SNP array, Axiom Arachis2, was designed using the approach which revealed 75% accuracy of calling SNPs from different tetraploid peanut genotypes. Using the method to simulate SNP variation in peanut, cotton, wheat, and strawberry, we show that models built with our parameter sets achieve above 98% accuracy in selecting true SNPs. Additionally, models built with simulated genotypes were able to select true SNPs at above 80% accuracy using real peanut data, demonstrating that our model can be used even if real data are not available to train the models. This work demonstrates an effective approach for calling highly reliable SNPs from polyploids using machine learning. A novel tool was developed for predicting true SNPs from sequence data, designated as SNP-ML (SNP-Machine Learning, pronounced \"snip mill\"), using the described models. SNP-ML additionally provides functionality to train new models not included in this study for customized use, designated SNP-MLer (SNP-Machine Learner, pronounced \"snip miller\"). SNP-ML is freely available for public use.