Inference of Synthetically Lethal Pairs of Genes Involved in Metastatic Processes via Non-Negative Matrix Tri-Factorization

Metastases formation dramatically represents the most widespread cause of cancer-related deaths; to date, however, few agents are specifically designed to contrast this phenomenon. In the last years, the concept of synthetic lethality (SL) for the design of new therapies has been effectively analysed and exploited. For the proper employment of this concept, a reliable computational identification of genetic SL pairs is required; massive use of laboratory-based methods is unfeasible due to the huge number of possible gene combinations. Non-Negative Matrix Tri-Factorization is emerging as an effective prediction method for inferring gene-to-gene relationships, and we hereby apply it to the inference of unknown SL gene pairs. The method is based on the decomposition of an association matrices in three non-negative matrices and their reconstruction. Here, the method is applied to data from SynLethDB, the most complete SL dataset, projected over metastatic genes. We show, by means of classical indicators, that he method is effective in inferring unseen pairs; we also demonstrate that the evaluation metric raises when we include in the prediction also other datasets describing gene features, that we properly include in the computation by means of additional matrices on the side of the main SL one. In this enhanced setting, we show that the method infers novel SL pairs, some of which independently supported by the literature.