Drug-Target Interaction Prediction Using Relative Multi-Head Self-Attention and Graph Attention Exchange Network

Abstract

Finding drug-target interactions will drastically limit the number of candidate medications that need to be looked for, making it the crucial initial stage in the drug discovery process. The number of biochemical experiments that can be undertaken is significantly constrained by the use of DTIs in the choice of potential drugs. They can also provide genuine insight on the health consequences and operating principles of drugs. Even now, research designs for finding drug-target interactions are nevertheless cost prohibitive, time-consuming, and challenging despite the availability of various biological techniques and elevated testing. As a result, numerous digital patterns have been created to broadly forecast possible drug-target relationships. This work proposes an Attention Exchange network of graph and encoder model for DTI (RMHSA_GAEN).   First, a GNN  receives pre-processed molecular graphs from SMILES from RDKit  of the drug compounds. Similarly encoder relying on a relative multi-head self-attention is used extract the feature vectors of the    protein sequences in the form of n-grams.   The proposed concept RMHSA_GAEN exchange of its individual attention weights to improve the feature correlation and  fine-tune the output from the graph network and Relative Multi Head Attention Transformer blocks. This model achieved accuracy 95.16% for the Human Dataset and 96.55% for the C.Elegans Dataset during testing on those datasets.