Abstract
Artificial neural networks in drug discovery have shown remarkable potential in various areas, including molecular similarity assessment and virtual screening. This study presents a novel multimodal Siamese neural network architecture. The aim was to join molecular electrostatic potential (MEP) images with the texture features derived from reduced density gradient (RDG) diagrams for enhanced molecular similarity prediction. On one side, the proposed model is combined with a convolutional neural network (CNN) for processing MEP visual information. This data is added to the multilayer perceptron (MLP) that extracts texture features from gray-level co-occurrence matrices (GLCM) computed from RDG diagrams. Both representations converge through a multimodal projector into a shared embedding space, which was trained using triplet loss to learn similarity and dissimilarity patterns. Limitations associated with the use of purely structural descriptors were overcome by incorporating non-covalent interaction information through RDG profiles, which enables the identification of bioisosteric relationships needed for rational drug design. Three datasets were used to evaluate the performance of the developed model: tyrosine kinase inhibitors (TKIs) targeting the mutant T315I BCR-ABL receptor for the treatment of chronic myeloid leukemia, acetylcholinesterase inhibitors (AChEIs) for Alzheimer’s disease therapy, and heterodimeric AChEI candidates for cross-validation. The visual and texture features of the Siamese architecture help in the capture of molecular similarities based on electrostatic and non-covalent interaction profiles. Therefore, the developed protocol offers a suitable approach in computational drug discovery, being a promising framework for virtual screening, drug repositioning, and the identification of novel therapeutic candidates.