Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Thesis
Deep learning for chromatin loop identification: comparing U-Net backbones with and without pre-training
Master of Science (M.S.), Drexel University
Dec 2025
DOI:
https://doi.org/10.17918/00011239
Abstract
Chromatin loops are essential three-dimensional genomic structures that regulate gene expression and influence disease risk. Detecting these loops from Hi-C contact maps is challenging due to data sparsity and scale. Building on the GILoop framework, we introduce two enhanced deep learning architectures: EU-Net, an optimized U-Net variant, and LoopNet, which integrates a pre-trained ResNet50 encoder with a custom decoder and classifier. Beyond model development, we systematically evaluate key factors such as patch size, resolution, and normalization strategies. Our results demonstrate that architectural refinements and transfer learning substantially improve performance, achieving more than a 2-fold increase in recall compared to the baseline GILoop model, while maintaining competitive precision. These findings provide a robust foundation for scalable and accurate chromatin loop prediction.
Metrics
13 File views/ downloads
18 Record Views
Details
- Title
- Deep learning for chromatin loop identification
- Creators
- Alexander F. Thoennes
- Contributors
- Ahmet Sacan (Advisor)Lei Wang (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Master of Science (M.S.)
- Publisher
- Drexel University
- Number of pages
- xix, 110 pages
- Resource Type
- Thesis
- Language
- English
- Academic Unit
- College of Computing and Informatics (2013-2026); Drexel University
- Other Identifier
- 991022148607304721