Fig. 3
From: AttCRISPR: a spacetime interpretable model for prediction of sgRNA on-target activity
The architecture of temporal domain method in AttCRISPR. The input of the method is embedding the sgRNA sequence Xe. Then Keys K, Values V and Queries Q are generated through a classic encoder-decoder structure which is needed by the temporal attention module. Next, the temporal attention module generates the first-order preference \(\tilde{A}\). Each of the row vectors in matrix \(\tilde{A}\) represents the base preference of sgRNA at the corresponding position, we use their dot product with the corresponding row vector in embedded Xe to build the score of the corresponding position. Hence, a full connection layer is used to weighted average them and achieve the temporal output yt