Figure 1
From: Network modeling of patients' biomolecular profiles for clinical phenotype/outcome prediction
Main logical steps of the P-Net algorithm. For simplicity the graph is shown with n = 6 patients. (1) Each patient is represented by a vector of m features (e.g. expression levels of the genes). Orange patients have the C phenotype; green patients do not show the C phenotype; violet patients are not labelled and our goal is to predict their label. We select the \(m{\prime} < m\) features most correlated with the phenotype C and we use them to construct a \(m{\prime} \times n\) matrix whose n columns represent the bio-molecular profiles of patients restricted to the \({m}^{{\prime} }\) selected features. (2) A W similarity matrix is constructed, where each element wij represents, e.g., the positive filtered Pearson correlation between the patients i and j; note that this matrix can be interpreted as the adjacency matrix of the graph of patients. To make the figure readable we show only the entries of patient p2. (3) The corresponding K kernel matrix is computed through e.g. 2-step Random Walk Kernel. (4) The K matrix is filtered and we remove all the edges with a weight lower than the selected threshold τ = 0.32. (5) A score function (e.g. Nearest Neighbour score) is finally used to compute a score for each patient and to rank them according to their scores.
