Deep Learning Strategies for Enhanced Molecular Docking and Virtual Screening

Over the last few years, machine learning (ML) and deep learning (DL) have been revolutionising the computer-aided drug discovery landscape. With the recent availability of the so-called ultra-large virtual libraries (libraries with up to billions of readily available virtual compounds), new ML and DL approaches have been developed to enable the exploration of these large chemical spaces, achieving promising results. Molecular docking is one of the most widely used computational methods for performing in silico screenings of virtual libraries. The two primary goals of molecular docking are to predict the correct binding pose of small molecules inside the binding pocket of a protein target and also estimate the binding affinity of the protein-ligand complex. In particular, DL methods have been applied in all aspects of protein-ligand molecular docking, from pose and binding affinity prediction to virtual screening campaigns, improving computational costs and accuracy. This chapter introduces the core aspects of the molecular docking methodology and some fundamental concepts of machine learning and deep learning. We also describe different types of molecular representations and DL architectures commonly employed in the field, such as convolutional and graph neural networks. Furthermore, we provide insights into potential applications by presenting related works from the scientific literature. Finally, we discuss the current limitations, challenges, and biases of DL applied to molecular docking.