Detecting Hybrid Incompatibilities with Machine Learning
When different species mate, hybrids are often inviable, sterile, or otherwise have decreased fitness. One cause of decreased hybrid fitness is hybrid incompatibilities: harmful epistatic interactions between genes inherited from the different parent species. The most well-understood hybrid incompatibilities are noticed due to having severe effects on hybrids, such as being fatal to all hybrid males, and the genes involved are pinpointed using extensive crosses. Subtler incompatibilities can be searched for by analyzing inheritance patterns in the genomes of hybrids. Previous studies suggest hundreds of these incompatibilities could exist between species. However, few large-scale analyses have been executed, and there are no standardized methods. Here, I explore a new method by developing neural networks to identify incompatibilities and estimate the effect they have on hybrid fitness. While the resulting neural networks are capable of detecting severe incompatibilities, they struggle with subtle incompatibilities and statistical power, like existing methods.