Challenges in Predicting Chromatin Accessibility Differences between Species

Area of Science:

• Genomics
• Computational Biology
• Evolutionary Biology

Background:

• Enhancers are key transcriptional regulatory elements driving phenotypic diversity.
• Rapid sequence evolution of enhancers despite functional conservation complicates cross-species functional prediction.
• Machine learning models for enhancer activity prediction have not been rigorously tested for cross-species quantitative differences.

Purpose of the Study:

• To evaluate the ability of machine learning models to predict quantitative differences in enhancer activity across orthologous regions in different species.
• To develop and apply a framework for assessing cross-species performance of enhancer activity prediction models.
• To investigate the impact of multi-species training data on model generalization and cross-species prediction.

Main Methods:

• Convolutional neural networks (CNNs) were trained on a regression task to predict chromatin accessibility (a proxy for enhancer activity) in the liver across five mammalian species.
• A novel framework was developed to evaluate the cross-species predictive performance of these CNN models.
• Model performance was assessed for both intra-species and inter-species prediction of chromatin accessibility differences.

Main Results:

• Training CNNs on multiple mammalian species improved model generalization to both training and held-out species.
• Models consistently demonstrated poor performance in predicting quantitative differences in chromatin accessibility between orthologous regions across species.
• Cross-species prediction accuracy for chromatin accessibility differences remained a significant challenge despite improvements in overall model generalization.

Conclusions:

• Multi-species training enhances the generalization of enhancer activity prediction models but does not fully resolve the challenge of predicting quantitative differences across species.
• Predicting evolutionary changes in enhancer activity and chromatin accessibility between species using current machine learning regression models remains difficult.
• Further development of computational frameworks is needed to accurately model the functional divergence of regulatory elements across evolutionary timescales.