Evaluation of object detection neural networks to find topologically associated domains in bulk Hi-C data
- 1University of Applied Sciences Mittelhessen
Abstract
Topologically associated domains (TADs) shape the three-dimensional structure of chromatin and contribute to gene regulation by arranging regulatory elements and genes in spatial proximity. This spatial proximity is captured by Hi-C, an experimental method, which quantifies contact frequencies between all pairs of genomic loci. In bulk Hi-C data, TADs can be found in the form of gapped, nested and partially overlapping squares centered on the main diagonal of the Hi-C matrix. We shift the problem of TAD identification into the domain of computer vision by detecting TADs in Hi-C in the form of squares in images. For this we use object detection, which aims to classify object instances in images. This study evaluates the applicability of the object detection neural networks YOLOv8 and Mask R-CNN on the identification of TADs in bulk Hi-C data. We build image-like data from Hi-C matrices, predict reference TADs with a state-of-the-art TAD caller, train the models and evaluate the detections. Our results confirm that YOLOv8 and Mask R-CNN learn to detect reference TADs, new TADs and overlapping TADs (although absent in training data). All three show increased ChIP-Seq signal of the TAD-associated factors CTCF, RAD21 and H3K4me3 on their boundaries, emphasizing their biological reliability. Comparisons with reference TADs indicate that Mask R-CNN detects more TADs with less precise positioning, while YOLOv8 detects fewer TADs with lower nesting complexity but improved positioning.
Related
- Identifying Putative Pathogenic Non-Coding Variants in Unresolved Rare Disease Patients Using Topologically Associated Domains
- Protected domains and cooperative invasion on weighted networks
- DNA damage at Lamina-Associated Domains triggers nuclear envelope reorganization and chromatin detachment to prevent nuclear envelope blebbing and genome instability.
- Assessing JULES across bioclimatic domains through a Pan-African flux tower evaluation: Insights into controls on gross primary productivity, ecosystem respiration, and evapotranspiration biases
- A comparison of artificial neural networks and partial least squares modelling for the rapid detection of the microbial spoilage of beef fillets based on Fourier transform infrared spectral fingerprints
BibTeX
@misc{philipp2026evaluation,
title = {{Evaluation of object detection neural networks to find topologically associated domains in bulk Hi-C data}},
author = {N. Philipp and C. L. Gruennagel and S. Janssen and C. Sigges},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.20.760189},
url = {https://www.biorxiv.org/content/10.64898/2026.10.20.760189v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Evaluation of object detection neural networks to find topologically associated domains in bulk Hi-C data
AU - N. Philipp
AU - C. L. Gruennagel
AU - S. Janssen
AU - C. Sigges
PY - 2026
DA - 2026/10/20
PB - bioRxiv
DO - 10.64898/2026.10.20.760189
UR - https://www.biorxiv.org/content/10.64898/2026.10.20.760189v1
AB - Topologically associated domains (TADs) shape the three-dimensional structure of chromatin and contribute to gene regulation by arranging regulatory elements and genes in spatial proximity. This spatial proximity is captured by Hi-C, an experimental method, which quantifies contact frequencies between all pairs of genomic loci. In bulk Hi-C data, TADs can be found in the form of gapped, nested and partially overlapping squares centered on the main diagonal of the Hi-C matrix. We shift the problem of TAD identification into the domain of computer vision by detecting TADs in Hi-C in the form of squares in images. For this we use object detection, which aims to classify object instances in images. This study evaluates the applicability of the object detection neural networks YOLOv8 and Mask R-CNN on the identification of TADs in bulk Hi-C data. We build image-like data from Hi-C matrices, predict reference TADs with a state-of-the-art TAD caller, train the models and evaluate the detections. Our results confirm that YOLOv8 and Mask R-CNN learn to detect reference TADs, new TADs and overlapping TADs (although absent in training data). All three show increased ChIP-Seq signal of the TAD-associated factors CTCF, RAD21 and H3K4me3 on their boundaries, emphasizing their biological reliability. Comparisons with reference TADs indicate that Mask R-CNN detects more TADs with less precise positioning, while YOLOv8 detects fewer TADs with lower nesting complexity but improved positioning.
ER -