Predicting Targets for Genome Editing with Long Short Term Memory Networks

Naturally occurring bacterial immune system can be engineered for use in the mammalian genome editing. To assist with the design of new editing systems, we developed and evaluated three data-driven predictors of DNA targets in the mouse and human genomes. Long Short Term Memory network models outperformed classifiers trained with Support Vector Machines and Random forest algorithms. The hold-out accuracy of the deep learning classifier reached 81.6% for the mouse genome and 82.5% for the human genome. We also demonstrated that classification accuracy improves when sequences surrounding a mammalian target site are incorporated into the input vector of the neural network, reaching an accuracy of 83% for both organisms.