DCGCN: Dual-Channel Graph Convolutional Network-Based Drug–Target Interaction Prediction Method with 3D Molecular Structure

Exploring drug–target interactions (DTIs) is crucial
for
drug discovery. Most existing methods for predicting DTIs rely solely
on the linear structures of molecules, such as SMILES or the amino
acid sequence. However, these linear features fail to reflect the
substructures of molecules or the relative positions of atoms. The
2D molecular structures, such as skeletal formulas or atom graphs,
also have limitations in fully reflecting the chemical structure of
molecules. To fully leverage the chemical structure of molecules,
this paper proposes DCGCN, a DTI prediction method based on 3D molecular
structure. DCGCN decomposes the 3D point cloud data of a molecule
into three components: atomic sequence, atomic connectivity, and a
distance map. From its connectivity and distance information, DCGCN
captures the relationships among atoms through a dual-channel graph
convolutional network. Furthermore, 1D convolutional layers are employed
to extract the chemical components with sequence information. Experimental
results on two public data sets demonstrate that DCGCN outperforms
several state-of-the-art DTI prediction methods, indicating that incorporating
the 3D structures of molecules can significantly improve DTI identification.