Challenges in predicting chromatin accessibility differences between species.

Differences in enhancer activity between species can help drive phenotypic diversity, yet enhancers often have conserved functions despite rapid sequence evolution, posing a challenge for quantifying their functional differences between species. Previous machine learning models have focused on the binary task of predicting differences in the presence of enhancers between species but have yet to demonstrate an ability to predict continuous differences in enhancer activity. Here, we trained convolutional neural networks on a regression task to predict chromatin accessibility-a proxy for enhancer activity-in the liver across five mammals, and we developed a novel framework to evaluate cross-species performance. We demonstrated that training on multiple species improves model generalization to both species used in training and held-out species. However, the models consistently achieved poor performance in predicting quantitative differences in accessibility between species at orthologous regions. Our study highlights the challenges in using regression models to predict chromatin accessibility changes between species. All data and code are available at http://daphne.compbio.cs.cmu.edu/files/azstephe/liver_regression_resource/ and https://figshare.com/projects/liverRegression/274293.