Describing conformational changes in biomolecules using molecular dynamics simulations requires defining an appropriate low-dimensional mathematical description of the system, referred to as a set of collective variables (CVs). No single CV design strategy is universally optimal; the choice should be guided by the biological question, the property of interest, the evaluation criterion, and the chosen sampling method. Here, we discuss the physical principles that should inform CV design and categorize existing approaches. We also evaluate the relationship between different types of CVs, the amount of data required to train them, and suitable enhanced sampling approaches. Finally, we outline practical guidelines for selecting CVs, helping practitioners match methodological choices to the underlying dynamical process and to the goals of their simulations.

The role of short strong hydrogen bonds (SSHBs) in ligand-target binding remains largely unexplored, thereby hindering a potentially important avenue in rational drug design. Here we investigate the interaction between the antituberculosis drug bedaquiline (Bq) and the mycobacterial ATP synthase to unravel the role of a specific hydrogen bond to a conserved acidic residue in the target affinity and specificity. Our ab initio molecular dynamics simulations reveal that this bond belongs to the SSHB category and accounts for a substantial fraction of the target binding free energy. We also demonstrate that the presence of an extra acidic residue, i.e., aspartic acid at position 32 (D32), found exclusively in mycobacteria, cooperatively enhances the HB strength, ensuring specificity for the mycobacterial target. Consistently, we show that the removal of D32 markedly weakens the affinity, leading to Bq resistance associated with mutations of D32 to nonacidic residues. By designing simple Bq analogs, we then explore the possibility to overcome the resistance and potentially broaden the Bq antimicrobial spectrum by making the SSHB independent of the presence of the extra acidic residue.

With the advent of AI-powered structure prediction, the scientific community is inching closer to solving protein folding. An unresolved enigma, however, is to accurately, reliably, and deterministically predict alternative conformational states that are crucial for the function of, e.g., transporters, receptors, or ion channels where conformational cycling is innately coupled to protein function. Accurately discovering and exploring all conformational states of membrane proteins has been challenging due to the need to retain atomistic detail while enhancing the sampling along interesting degrees of freedom. The challenges include but are not limited to finding which degrees of freedom are relevant, how to accelerate the sampling along them, and then quantifying the populations of each micro- and macrostate. In this work, we present a methodology that finds relevant degrees of freedom by combining evolution and physics through machine learning and apply it to the conformational sampling of the beta 2 adrenergic receptor. In addition to predicting new conformations that are beyond the training set, we have computed free energy surfaces associated with the protein's conformational landscape. We then show that the methodology is able to quantitatively predict the effect of an array of ligands on the beta 2 adrenergic receptor activation through the discovery of new metastable states not present in the training set. Lastly, we also stake out the structural determinants of activation and inactivation pathway signaling through different ligands and compare them to functional experiments to validate our methodology and potentially gain further insights into the activation mechanism of the beta 2 adrenergic receptor.