Analyzing Genomic Features with Predictive Chromatin Interaction Models: A Comprehensive Evaluation

Enhancer-Promoter (EP) interactions reflected by Hi-C technology are crucial to understanding genomic functions. Particularly, identifying `unique' genomic features that are characteristically important in a specific cell line can further our current understanding of the mechanisms that drive cell differentiation, tissue development, and disease progression. However, classic prediction models such as TargetFinder provide little valuable insight towards the large disparity between important genomic features across different cell lines. To comprehensively approach this question, herein we first evaluated seven classifiers to predict EP interactions using high-resolution Hi-C maps of genome loci across six classic cell lines, surpassing TargetFinder in all benchmark metrics. We then evaluated the model's predictive performance with features provided by seven feature selection methods from the embedded, wrapper and filter categories. Moreover, groups of features were aggregated from the results of two or more feature methods and analyzed based on the model's performance. Finally, we examined the distinguishing features across six cell lines. Our study suggests the existence of `unique' genomic features that are especially predictive of EP interactions only in specific cell lines.