Predicting miRNA-Disease Association Based on Neural Inductive Matrix Completion with Graph Autoencoders and Self-Attention Mechanism

. 2022 Jan 2;12(1):64. doi: 10.3390/biom12010064

Algorithm 1 NIMGSA Algorithm

Input:
initial association matrix Y, miRNA similarity matrix $S_{m}$ , disease similarity matrix $S_{d}$
Output:
score matrix F
1:
Compute the adjacency matrix of miRNA graph $A_{m}$ and disease graph $A_{d}$ via Equations (21) and (22) respectively
2:
repeat
3:
Compute Query: $Z_{m} = Enc (A_{m}, Y)$ // Equation (19)
4:
Compute Key: $Z_{d} = Enc (A_{d}, Y^{T})$ // Equation (20)
5:
$F_{m} = Dec (A_{m}, Z_{m})$ // Equation (25)
6:
$F_{d} = Dec (A_{m}, Z_{d})$ // Equation (26)
7:
Compute Value: $F = α F_{m} + (1 - α) F_{d}^{T}$ // Equation (28)
8:
Compute attention score as Equation (29)
9:
Train GAEm and GAEd through optimizing Equation (32) // Update W (i.e., parameters of neural networks) in Equation (32) by Adam optimizer
10:
until Convergence
11:
return F