Fig. 2

Overview of DDINMF. DDINMF contains a training phase and a predicting phase. (1) In its training phase, the adjacent matrix A is first decomposed into a basis (community) matrix and a latent (encoding) feature matrix by A ≈ W × H. Then the relationship between the input feature matrix F and the latent feature matrix H is modeled by a regression (H^T) = F × B. (2) In the predicting phase, the learned regression coefficient B firstly maps the input feature matrix F _x of n newly given drugs into their latent feature matrix by \( {\mathbf{H}}_x^T={\mathbf{F}}_x\times \mathbf{B} \). Then the mapped latent feature matrix of F _x is used to generate the predicted interactions between the new drugs and the known drugs by A _x  = (WH _x )^T