Fast Structure-Based De Novo Drug Design with Molecular Encoders and Computational Intelligence

Drug design is the first step of the long and expensive process, named drug development, aimed at launching a new drug on the market. Due to the high cost and time required by experimental techniques, in silico methods represent a valuable tool to help speed up and make the development more efficient. In this work, we propose EVOMOL-FLOW, an approach that combines evolutionary computation and deep learning models, specifically tailored to this problem. A multi-objective optimization algorithm – with novel evolutionary operators – is used to design novel drug candidates, maximizing their binding affinity to the target while minimizing the difficulty of their synthesis in the laboratory. To reduce the bottlenecks of the fitness function evaluation, we take advantage of DrugCLIP, a molecular encoder that maps the chemical structures of the ligand and the pocket into a latent space, and determines the binding affinity by measuring the cosine distance between the embeddings. This approach reduces the overall running time by orders of magnitude, paving the way to large-scale and more accurate evolution of candidate drugs. As a case study, we applied EVOMOL-FLOW to the design of putative molecules targeting Vascular Endothelial Growth Factor Receptor (VEGFR), a membrane protein responsible for vascularisation in solid tumors. To show the effectiveness of the approach, we further validate the produced compounds by means of molecular dynamics and docking experiments.