Publication Date
Fall 2025
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Engineering
Advisor
Carlos Rojas; Bernardo Flores; Mahima Suresh
Abstract
High-throughput chromosome conformation capture (Hi-C) enables interpretation of three-dimensional genome organization. Traditional Hi-C aggregates millions of cells to form symmetric contact maps, whereas single-cell Hi-C (scHi-C) captures per-cell interaction maps that reveal cell-to-cell heterogeneity, but at much sparser coverage. Because generating high-resolution Hi-C maps is costly, learning low- to high-resolution mappings with computer vision models is a practical alternative. This study compares the standard Hi-C Plus model with a FiLM-augmented Hi-C Plus variant that applies feature-wise linear modulation to improve reconstruction quality. Using bulk and single-cell datasets, we evaluate performance using Pearson and Spearman correlation, mean squared error (MSE), and structural similarity (SSIM). Across conditions, FiLM conditioning improves reconstruction quality relative to Hi-C Plus, with the largest gains observed on sparse single-cell data.
Recommended Citation
Shnayder, Arielle, "Super-Resolving Hi-C: Benchmarking Hicplus and A Film-Conditioned Resnet on Bulk and Single-Cell Data" (2025). Master's Theses. 5714.
DOI: https://doi.org/10.31979/etd.sxay-3cuj
https://scholarworks.sjsu.edu/etd_theses/5714
