MaZhang19, Integrate multi-omics data with biological interaction networks using Multi-view Factorization AutoEncoder (MAE)

 

Integrate multi-omics data with biological interaction networks using

Multi-view Factorization AutoEncoder (MAE)

Tianle Ma, Aidong Zhang, 2019, BMC genomics

My understanding the the deep learning its is not biological network. But biological nextwork adjacency matrix x inferred network matrix is used as a regularization term. 

Ma-Zhang19 used feature interaction network -> Laplacian matrix of graph. 

Ma-Zhang19 used encoder-decoder because this is similar to Matrix factorization, which need to keep the same dimensions of matrices, based on Hong's understanding. 

Ma suggested that autoencoder basically works like PCA, scale input matrix into the orthological ones. So, autodencoder theoretically factorize matrix based on this rank, though in practice, the hidden variable dimension is a hyper-parameter decided by trial-and-errors. Ma suggested that network-based learning can be used after the encoder layer.  

It seems that auto-encoder of the network input might be achieved by the "driver nodes".