Molecular self-organization driven by concerted many-body interactions produces the ordered structures that define both inanimate and living matter. Here we present an autonomous path sampling algorithm that integrates deep learning and transition path theory to discover the mechanism of molecular self-organization phenomena. The algorithm uses the outcome of newly initiated trajectories to construct, validate and—if needed—update quantitative mechanistic models. Closing the learning cycle, the models guide the sampling to enhance the sampling of rare assembly events. Symbolic regression condenses the learned mechanism into a human-interpretable form in terms of relevant physical observables. Applied to ion association in solution, gas-hydrate crystal formation, polymer folding and membrane-protein assembly, we capture the many-body solvent motions governing the assembly process, identify the variables of classical nucleation theory, uncover the folding mechanism at different levels of resolution and reveal competing assembly pathways. The mechanistic descriptions are transferable across thermodynamic states and chemical space.
Transition path sampling is a powerful tool in the study of rare events. Shooting trial trajectories from configurations along existing transition paths proved particularly efficient in the sampling of reactive trajectories. However, most shooting attempts tend not to result in transition paths, in particular in cases where the transition dynamics has diffusive character. To overcome the resulting efficiency problem, we developed an algorithm for “shooting from the top.” We first define a shooting range through which all paths have to pass and then shoot off trial trajectories only from within this range. For a well chosen shooting range, nearly every shot is successful, resulting in an accepted transition path. To deal with multiple mechanisms, weighted shooting ranges can be used. To cope with the problem of unsuitably placed shooting ranges, we developed an algorithm that iteratively improves the location of the shooting range. The transition path sampling procedure is illustrated for models of diffusive and Langevin dynamics. The method should be particularly useful in cases where the transition paths are long so that only relatively few shots are possible, yet reasonable order parameters are known.
21 November 2023 in Journal of Chemical Theory and Computation
Molecular dynamics is a powerful tool for studying the thermodynamics and kinetics of complex molecular events. However, these simulations can rarely sample the required time scales in practice. Transition path sampling overcomes this limitation by collecting unbiased trajectories and capturing the relevant events. Moreover, the integration of machine learning can boost the sampling while simultaneously learning a quantitative representation of the mechanism. Still, the resulting trajectories are by construction non-Boltzmann-distributed, preventing the calculation of free energies and rates. We developed an algorithm to approximate the equilibrium path ensemble from machine-learning-guided path sampling data. At the same time, our algorithm provides efficient sampling, mechanism, free energy, and rates of rare molecular events at a very moderate computational cost. We tested the method on the folding of the mini-protein chignolin. Our algorithm is straightforward and data-efficient, opening the door to applications in many challenging molecular systems.
13 July 2021 in The Journal of Physical Chemistry B
During infection the SARS-CoV-2 virus fuses its viral envelope with cellular membranes of its human host. The viral spike (S) protein mediates both the initial contact with the host cell and the subsequent membrane fusion. Proteolytic cleavage of S at the S2′ site exposes its fusion peptide (FP) as the new N-terminus. By binding to the host membrane, the FP anchors the virus to the host cell. The reorganization of S2 between virus and host then pulls the two membranes together. Here we use molecular dynamics (MD) simulations to study the two core functions of the SARS-CoV-2 FP: to attach quickly to cellular membranes and to form an anchor strong enough to withstand the mechanical force during membrane fusion. In eight 10 μs long MD simulations of FP in proximity to endosomal and plasma membranes, we find that FP binds spontaneously to the membranes and that binding proceeds predominantly by insertion of two short amphipathic helices into the membrane interface. Connected via a flexible linker, the two helices can bind the membrane independently, yet binding of one promotes the binding of the other by tethering it close to the target membrane. By simulating mechanical pulling forces acting on the C-terminus of the FP, we then show that the bound FP can bear forces up to 250 pN before detaching from the membrane. This detachment force is more than 10-fold higher than an estimate of the force required to pull host and viral membranes together for fusion. We identify a fully conserved disulfide bridge in the FP as a major factor for the high mechanical stability of the FP membrane anchor. We conclude, first, that the sequential binding of two short amphipathic helices allows the SARS-CoV-2 FP to insert quickly into the target membrane, before the virion is swept away after shedding the S1 domain connecting it to the host cell receptor. Second, we conclude that the double attachment and the conserved disulfide bridge establish the strong anchoring required for subsequent membrane fusion. Multiple distinct membrane-anchoring elements ensure high avidity and high mechanical strength of FP–membrane binding.
P-glycoprotein (Pgp) is a prototypical ATP-binding cassette (ABC) transporter of great biological and clinical significance.Pgp confers cancer multidrug resistance and mediates the bioavailability and pharmacokinetics of many drugs (Juliano and Ling, 1976; Ueda et al., 1986; Sharom, 2011). Decades of structural and biochemical studies have provided insights into how Pgp binds diverse compounds (Loo and Clarke, 2000; Loo et al., 2009; Aller et al., 2009; Alam et al., 2019; Nosol et al., 2020; Chufan et al., 2015), but how they are translocated through the membrane has remained elusive. Here, we covalently attached a cyclic substrate to discrete sites of Pgp and determined multiple complex structures in inward- and outward-facing states by cryoEM. In conjunction with molecular dynamics simulations, our structures trace the substrate passage across the membrane and identify conformational changes in transmembrane helix 1 (TM1) as regulators of substrate transport. In mid-transport conformations, TM1 breaks at glycine 72. Mutation of this residue significantly impairs drug transport of Pgp in vivo, corroborating the importance of its regulatory role. Importantly, our data suggest that the cyclic substrate can exit Pgp without the requirement of a wide-open outward-facing conformation, diverting from the common efflux model for Pgp and other ABC exporters. The substrate transport mechanism of Pgp revealed here pinpoints critical targets for future drug discovery studies of this medically relevant system.
26 August 2025 in Proceedings of the National Academy of Sciences
Stromal interaction molecule 1 (STIM1) is a Ca 2+ -sensing protein in the endoplasmic reticulum (ER) membrane. The depletion of ER Ca 2+ stores induces a large conformational transition of the cytosolic STIM1 C-terminus, initiated by the dimerization of the transmembrane (TM) domain. We use the AI-guided transition path sampling algorithm aimmd to extensively sample the dimerization of STIM1-TM helices in an ER-mimicking lipid bilayer. In nearly 0.5 ms of all-atom molecular dynamics simulations without bias potentials, we harvest over 170 transition paths, each about 1.2 μs long on average. We find that STIM1 dimerizes into three distinct and coexisting configurations, which reconciles conflicting results from earlier crosslinking studies. The dominant X-shaped bound state centers around contacts supported by the SxxxG TM interfacial motif. Mutating residues in this contact interface allows us to tune the STIM1-dimerization propensity in fluorescence experiments. From the trained model of the committor probability of dimerization, we identify the transition state ensemble for TM-helix dimerization. At the transition state, interhelical contacts in the luminal halves of the two monomers dominate, which likely enables the luminal Ca 2+ -sensing domain in STIM1 to condition the dimerization of the TM helices. Our work demonstrates the unique power of AI-guided simulations to sample rare and slow molecular transitions and to produce detailed atomistic insight into the mechanism of STIM1 TM-helix dimerization as a key step in ER Ca 2+ -sensing.