ResNet Combined with Attention Mechanism for Genomic Deletion Variant Prediction
Abstract
In genetics and medical practice, structural variants (SV) in the genome are thought to be the root cause of numerous diseases, particularly genetic diseases. Accurate structural variant prediction is the foundation for identifying and screening pathogenic variants and performing medication genomics analysis, which is a challenging task. However, data in the field of genomics is typically massive, high-dimensional, and serialized, and existing variant prediction tools are affected by the range and type of variants, resulting in less accurate results. As a result, an effective method for predicting structural variation is critical. In this paper, a variation prediction model DEL-RESSP based on ResNet and attention mechanism is proposed for predicting deletion structural variants. To begin, the deletion variant feature information is derived from the three alignment data of read depth, split read pair, and discordant read pair, and the comparison data is transformed into artificial images by encoding to provide reliable input for the subsequent network models. Second, attention mechanisms are combined based on convolutional networks to improve image sensitivity to local information to improve prediction accuracy. Three SV prediction tools, CNVnator, BreakDancer, and Pindel, were used in this study to test the predictive effectiveness of DEL-RESSP in predicting large-scale deletion variants. The results show that DEL-RESSP can predict deletion variants with 96.93% accuracy, which is a 5-10% improvement over combining only a single strategy, as well as a comparison to existing deep learning methods. DEL-RESSP fully utilizes deep learning in image processing, providing some reference value in subsequent variant analysis and gene function annotation. Part of the classification model code used in this paper can be found on https://github.com/JQ1209/DEL-RESSP.
