Multi-filter based signed heterogeneous graph convolutional networks for predicting activating/inhibiting drug-target interactions

Abstract

The prediction of mechanisms within drug-target interactions (DTIs) can boost the drug discovery process, which has traditionally relied on time-consuming and expensive laboratory experiments. Despite much more attention has been paid to predicting DTIs, but few studies focused on their activating/inhibiting mechanisms. In this work, we model DTIs on signed heterogeneous networks, through categorizing activating/inhibiting DTIs into signed links, and accordingly introducing the coherence/incoherence between drugs on a common target to construct signed drug-drug links. We propose a multi-filter based signed heterogeneous graph convolutional network (MFSHGCN) for drugs and targets embedding, via employing dual filters on both the signed drug-drug sub-graph and the signed DTI sub-graph to converge the spectral information from positive and negative edges. We further put forward an end-to-end framework to predict activation and inhibition within DTIs. The comparison results demonstrate the introduction of coherence/incoherence of drug pairs and the design of our multi-filter system can effectively improve the prediction metrics, even without relying on rich node information and interactions from drug pairs or target pairs. Case studies on breast cancer and lung cancer confirm the model's feasibility.