ProtContext-DTI: Protein Contextual Representation Using Masked Language Modeling in Drug Target Interaction Prediction

Comprehending the strength of interaction between drugs and their molecular targets holds significant importance in advancing drug development processes. This paper introduces a novel approach for predicting drug-target affinity (DTA), responding to the increasing focus on DTA in recent research. Recently, DTA has gained increased attention. The methods presented are categorized into ligand-based and similarity-based approaches. The first category feeds the protein sequence and drug molecules as input to learn feature representation. The similarity-based methods represent a protein (molecule) by a vector, which shows its similarity to all other proteins (drugs) in the datasets. These vectors are subsequently used as inputs for the model. The paper proposes a unified framework that integrates both ligand-based and similarity-based perspectives. A novel protein representation is introduced, employing a protein language model to learn the probability distribution for each amino acid based on its neighboring region. This encoding captures contextual information for each amino acid in its corresponding probability distribution. Then, this new representation is fed into a modified differentiable Smith-Waterman algorithm to compute the similarity matrix between proteins. Finally, the proposed architecture integrates the drug molecule, protein sequence, and the generated similarity matrix for predicting binding affinity. To assess the proposed method, it is applied to two well-known datasets: Davis and KIBA. Extensive experiments have been conducted to demonstrate its efficacy compared to state-of-the-art approaches.