Enhanced Sampling Simulations of RNA–Peptide Binding Using Deep Learning Collective Variables

Enhanced sampling (ES) simulations of biomolecular recognition, such as binding small molecules to proteins and nucleic acid targets, protein-protein association, and protein-nucleic acid interactions, have gained significant attention in the simulation community because of their ability to sample long-time scale processes. However, a key challenge in implementing collective variable (CV)-based enhanced sampling methods is the selection of appropriate CVs that can distinguish the system's metastable states and, when biased, can effectively sample these states. This challenge is particularly acute when the binding of a flexible molecule to a conformationally rich host molecule is simulated, such as the binding of a peptide to an RNA. In such cases, a large number of CVs are required to capture the conformations of both the host and the guest as well as the binding process. Using such a large number of descriptors is impractical in any enhanced sampling simulation method. In our work, we used the recently developed deep targeted discriminant analysis (Deep-TDA) method to design CVs to study the binding of a cyclic peptide, L22, to a TAR RNA of HIV, which is a prototypical system. The Deep-TDA CV, obtained from a nonlinear combination of important contact pairs between the L22 peptide and the host RNA backbone atoms, along with the RNA apical loop RMSD as the second CV were used in the on-the-fly probability-based enhanced sampling (OPES) simulation to sample the reversible binding and unbinding of the L22 peptide to the TAR RNA target. The OPES simulation delineated the mechanism of peptide binding and unbinding to and from the RNA and enabled the calculation of the underlying free energy landscape. Our results demonstrate the potential of the Deep-TDA method for designing CVs to study complex biomolecular recognition processes.