Poster #5: Kuanyang Fang

A Sequence-Based Framework for Predicting the Evolutionary Dynamics of Enhancer-Associated Liver CTCF Binding across Mammals

Kuanyang Fang, BS (1); Alexander S. Okamoto, PhD (1); Junjie Ma, PhD (1); Irene M. Kaplow, PhD (1,2). (1) Department of Biological Sciences, Carnegie Mellon University, Pittsburgh, PA, USA; (2) Ray and Stephanie Lane Computational Biology Department, Carnegie Mellon University, Pittsburgh, PA, USA.

CCCTC-binding factor (CTCF) is an 11-zinc-finger DNA-binding protein that contributes to context-dependent gene regulation and higher-order genome organization. Mammalian CTCF occupancy includes conserved and lineage-specific sites, yet experimental liver CTCF ChIP-seq data remain limited to a few species. Sequence-based prediction could therefore characterize enhancer-associated liver CTCF-binding dynamics across a broad phylogenetic scale. We developed a sequence-based framework to test whether local DNA features learned in the house mouse generalize across species and support prediction across hundreds of mammalian genomes with varying assembly quality. We trained a convolutional neural network on 500-bp sequences from accessible mouse liver candidate regulatory regions using chromosome-partitioned training, validation, and test sets. We evaluated within-mouse performance; transfer to species not used in training, including rat, human, macaque, and cow; cross-species binding-difference performance; specificity relative to other liver transcription factors and tissues; promoter-versus-enhancer performance; and CTCF motif-stratified performance. The preliminary mouse-trained model performed strongly within mouse (AUROC > 0.98; AUPRC > 0.97) and across four additional species (AUROC > 0.95; AUPRC > 0.91). It distinguished CTCF-binding from four other liver transcription factors (AUROC > 0.94) and performed strongly for binding shared across tissues, in promoters and enhancers, and within motif-present and motif-absent strata (AUROC > 0.92). Performance decreased but was still strong for rat-mouse orthologs with different binding labels (AUROC = 0.83) and was lower for liver-specific versus other-tissue-specific binding (AUROC = 0.56-0.82) and motif-absent bound versus motif-present unbound regions (AUROC = 0.66), likely because much of CTCF-binding is shared across tissues and at this motif. We will next train multispecies models and anticipate that they will learn a shared regulatory code that yields stronger performance. We will use the best to predict across more than 500 mammalian genomes to study CTCF-binding dynamics and associate changes in binding with phenotypic variation across species.