A Neural Network Approach for the Identification of Pathways in Molecular Dynamics Simulations of Ligand Binding

Motivation: Understanding the process of ligand-protein recognition is important to unveil biological mechanisms and to inform drug design and discovery. Enhanced sampling molecular dynamics (MD) methods can be used to simulate the ligand-binding process, obtaining information regarding its thermodynamics and kinetics (Limongelli V, 2020). The extensive sampling of the binding/unbinding events, through several simulation replicas or with a single simulation that describes several re-crossing events, creates the need for tools able to analyze a huge amount of data and to provide a clear picture of the sampled pathways. Methods: We propose that binding events can be analysed using Self-Organizing Maps (SOMs), a type of artificial neural network useful for effective identification of patterns in the data (Kohonen T 2013). SOMs have already been used successfully for the analysis of MD simulations (Motta S et al. 2021). We designed, implemented and tested a SOM-based tool to build and analyze models of ligand binding pathways sampled along a simulation or in different replicas. The tool is written in R and can be used to perform a geometric clustering of the trajectories under study to obtain an overview of the conformations sampled during the simulation; to trace the sampled pathways to recover the temporal evolution of the system on the SOM; to build a network model that represents the pathways in a clear way using a transition matrix built from the SOM neurons. Results: We tested the tool on different systems to demonstrate its general applicability and good performance. For this reason, the study-cases differ in the choice of MD methods used to investigate the ligand binding, including steered and metadynamic simulations of the THS-020 ligand binding to the HIF-2α protein (Callea L et al. 2021) and the infrequent metadynamic simulations of the GC7 ligand unbinding from human deoxyhypusine synthase (D’Agostino M et al. 2020). For both systems the tool was able not only to identify the preferred unbinding pathway but also to detect the position of the energetic barrier for the transition.