Fig. 1

Two categories of deep learning models are used in sgRNA related tasks. a Model work in the spatial domain. In the spatial domain, the base sequence is encoded into a binary matrix (or a binary image). Since convolution has great advantages in extracting spatial features, CNN is an excellent tool in the spatial domain. b Model work in the temporal domain. In the temporal domain, the base sequence (represented by the binary matrix) is embedded into a sequence of high-dimensional vectors, in which the RNN performs better. In addition, we note that the last layers of these neural networks are usually full connection structures (not necessarily), which greatly increases the difficulty of understanding the decisions of these models