Poster #42 - My Nguyen

Improving prediction of 3D chromatin structures (Topologically Associating Domains)

My Nguyen, BS, Department of Biostatistics, SOPH, VCU Shaojun Tang, , PhD, Department of Biostatistics, SOPH, VCU J. Chuck Harrell, PhD, Department of Pathology, SOM, VCU Mikhail G. Dozmorov, PhD, Department of Biostatistics, SOPH, VCU

The human genome can be divided into different 3D topologically associating domains (TADs). DNA regions within a TAD interact with each other more often than with regions outside the TAD. TAD boundary disruption is associated with a wide range of diseases such as cancer, neurological disorders, and development. Numerous methods have been developed to detect TAD boundaries using boundary probability vectors derived from chromatin contact maps and/or genomic data. However, these methods are largely limited by the resolution of Hi-C data, which typically ranges from 1 Kb to 100 Kb. In contrast, functional DNA loci span a much wider range-from a few hundred base pairs to over 50 Kb-with a median size of 10.5 Kb. To improve high-resolution boundary detection, we have hypothesis that the patterns of epigenetic signals associated with regions in proximity of TAD boundaries, which contain distinctive information that can serve as embeddings for genomic regions. These embeddings, along with their positional relationships, can be effectively modeled using deep learning to achieve more precise boundary prediction. To leverage this, we propose a transformer-based model that takes as input neighboring regions of transcriptional and histone modification signals centered around candidate boundaries. We also evaluate and compare the performance of other deep learning architectures, including feedforward neural networks, convolutional neural networks (CNNs), and bidirectional long short-term memory (BiLSTM) networks, using the same epigenetic input features. All models perform high accuracy in detecting TAD boundary, in which the transformer having higher performance compared to other models. The epigenetic regions surrounding TAD boundaries provides a strong predictive signal, enabling these models to accurately detect boundaries. These findings highlight the need to further study of deep learning approaches in refining the epigenomic language of TAD domains and its complex 3D genome organization.