Table 1 The devised piRNA–mRNA binding deep learning network structure
From: Identifying piRNA targets on mRNAs in C. elegans using a deep multi-head attention network
Sub-Networks | piRNA feature | mRNA feature |
|---|---|---|
I: Motif feature extraction sub-network | Conv (size \(=\) 5) * 128 | Conv (size \(=\) 5) * 128 |
Batch-normalization | Batch-normalization | |
PReLU | PReLU | |
SE block (P) | SE block (M) | |
II: Multi-head attentive binding recognition sub-network | 16-head attention layer (output: H) | |
Add Residual (K = H+M) | ||
Layer-normalization | ||
FC layer (hidden layer size = 128*4) | ||
PReLU | ||
FC layer (hidden layer size = 128) (L) | ||
Add Residual (K+L) | ||
Layer-normalization | ||
III: Classification sub-network | Flatten | |
FC layer (hidden layer size = 32*31) | ||
Batch-normalization | ||
PReLU | ||
FC layer (hidden layer size = 8*31) | ||
Batch-normalization | ||
PReLU | ||
Softmax layer | ||