Challenges in predicting chromatin accessibility differences between species.

Area of Science:

• Genomics and Evolutionary Biology
• Computational Biology and Bioinformatics

Background:

• Enhancer activity differences contribute to phenotypic diversity across species.
• Conserved enhancer functions contrast with rapid sequence evolution, complicating functional difference quantification.
• Previous machine learning models focused on binary enhancer presence prediction, not continuous activity levels.

Purpose of the Study:

• To train convolutional neural networks for predicting continuous differences in enhancer activity across species using chromatin accessibility as a proxy.
• To develop and evaluate a framework for assessing cross-species model performance in predicting quantitative enhancer activity.
• To investigate the efficacy of multi-species training for improving model generalization.

Main Methods:

• Convolutional neural networks were trained on a regression task to predict liver chromatin accessibility in five mammalian species.
• A novel framework was implemented to evaluate cross-species predictive performance.
• Model generalization was assessed using both in-training and held-out species.

Main Results:

• Training models on multiple species enhanced generalization to both familiar and novel species.
• Despite improvements in generalization, models consistently performed poorly in predicting quantitative differences in chromatin accessibility between species at orthologous regions.
• The study identified significant challenges in applying regression models for predicting inter-species chromatin accessibility changes.

Conclusions:

• Multi-species training can improve the generalization of machine learning models for predicting enhancer activity.
• Predicting quantitative differences in enhancer activity across species remains a significant challenge.
• Further research is needed to develop more robust methods for quantifying functional enhancer divergence.