A Neural Network Approach for Anomaly Detection in Genomic Signals

Structural variants (SVs) are observed differences between the sequenced genome of an individual as compared to a reference genome for that species. These differences include deletions, inversions, insertions, and duplications. Since some variations are associated with certain diseases, our work focuses on developing methods to detect such genomic anomalies. Current DNA sequencing methods may be costly and existing SV-detection techniques often rely on high quality data. We present a deep learning method to identify deletions in DNA based on genomic information of related individuals. In this paper, we implement neural networks to predict SVs as a means to reduce the false positive rates of existing methods. A neural network - a sequence of linear and nonlinear transformations - takes in training data and uses that information to learn how to classify corresponding test data. Our preliminary model incorporates the observed genomic information of two parents and an offspring to predict locations of SVs in the genome of the child. We also investigate the performance of this model under different neural network architectures using various performance metrics. With these limited features and low-quality data, we propose a generalization of our model that allows for the simultaneous prediction of SVs in all three individuals.