# Graph Neural Networks for Double-Strand DNA Breaks Prediction

**Authors:** XU Wang, Huan Zhao, Weiwei TU, Hao Li, Yu Sun, Xiaochen Bo  
**arXiv:** [2201.01855](https://arxiv.org/abs/2201.01855)  
**Date:** 2022-01-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.01855v1/

## Key points

- Double-strand DNA breaks (DSBs) can cause abnormal chromosomal rearrangements
- Detecting DSBs using high-throughput experiments is costly and technically challenging
- GraphDSB is a graph neural network-based method that predicts DSBs using DNA sequence features and chromosome structure information
- The model incorporates Jumping Knowledge architecture and several effective structural encoding methods to improve its expression ability
- GNNExplainer is used to analyze the contribution of node features and topology to DSB prediction, revealing that 5-mer DNA sequence features and two chromatin interaction modes have a high contribution
- Compared to other low-cost, high-efficiency DSB prediction methods such as Random Forest-based models, GraphDSB leverages chromosome structure information resulting in improved accuracy
- Studies show that GraphDSB outperforms other GNN variants on NHEK and K562 datasets
- This study proposes an effective method for predicting DSBs using graph neural networks based on DNA sequence features and chromosome structure information which improves accuracy compared to other low-cost methods while reducing technical challenges associated with high-throughput experiments
- Future work could explore how this approach could be applied in clinical settings or extended beyond human cells to other organisms

## Summary

Double-strand DNA breaks (DSBs) are a type of DNA damage that can lead to abnormal chromosomal rearrangements. Detecting DSBs using high-throughput experiments is costly and technically challenging. To address this issue, the authors propose GraphDSB, a graph neural network-based method that predicts DSBs using DNA sequence features and chromosome structure information. The model incorporates Jumping Knowledge architecture and several effective structural encoding methods to improve its expression ability. Ablation studies demonstrate the effectiveness of the designed components in the proposed GraphDSB framework. Additionally, GNNExplainer is used to analyze the contribution of node features and topology to DSB prediction, revealing that 5-mer DNA sequence features and two chromatin interaction modes have a high contribution. Compared to other low-cost, high-efficiency DSB prediction methods such as Random Forest-based models, GraphDSB leverages chromosome structure information when predicting DSBs which has been shown to have a correlation with DSBs resulting in improved accuracy. Studies show that GraphDSB outperforms other GNN variants on NHEK and K562 datasets. In conclusion, this study proposes an effective method for predicting DSBs using graph neural networks based on DNA sequence features and chromosome structure information which improves accuracy compared to other low-cost methods while reducing technical challenges associated with high-throughput experiments. Future work could explore how this approach could be applied in clinical settings or extended beyond human cells to other organisms.