Fig. 2
From: Influence network model uncovers relations between biological processes and mutational signatures
Workflow of GeneSigNet. A Input matrix X of node activities constructed by concatenating gene expression values (for genes) and signature exposures (for MutStates) across p samples (patients). B Given the input matrix X, we infer a network of n nodes. For each node, Sparse Partial Correlation Selection (SPCS) is used to simultaneously estimate the weights of incoming effects from the other \(n-1\) nodes. C For each bidirectional edge (j, f), the residual vectors \(r_j\) and \(r_f\) corresponding to nodes j and f are obtained by removing effects of the \(n-2\) nodes other than the two nodes of the considered edge. The non-zero weight \(w_{kf}\) obtained by SPCS denotes the strength of the confounding effect on node f coming from node k. The direction of a influence effect between the pair of nodes is determined based on the partial higher moment statistics, skewness and kurtosis of residuals \(r_j\) and \(r_f\). If both moments support the same direction with the heavier partial correlation weight (see Additional file 1: Equation S5), then the edge corresponding to the opposite direction is removed, otherwise, both edges remain in the network. D Edge weights are normalized using an alternative scaling algorithm, and the final weighted-directed network is obtained as the output