GUR LABORATORY

In Silico Pulling Experiments Aligned with High-Speed Atomic Force Microscopy Loading Rates

In silico single-molecule pulling experiments are typically carried out using steered molecular dynamics (SMD) simulation, a technique in which a “dummy atom” is attached to a virtual spring and pulled along a predefined vector at a constant velocity or force. However, computational limitations have historically restricted SMD simulations to pulling velocities of 2–100 Å/ns, with corresponding loading rates (pulling velocities × spring constant) of 200–1,000,000 pN/ns. These values can exceed high-speed atomic force microscopy (hAFM) conditions by ten orders of magnitude, leaving critical questions about the effects of extremely fast pulling unresolved—particularly regarding work irreversibility, deviations from free energy changes, and rupture forces.

To bridge this gap, my lab has conducted a large-scale series of all-atom SMD simulations—totaling 150 µs of simulation time and consuming 100 million core hours—at much lower pulling velocities (0.1 Å/ns) [1,2,3] and loading rates (1 pN/ns) [1], and we are actively extending these simulations beyond our published work. These parameters closely match those used in hAFM, providing an unprecedented view into an in silico regime that was previously inaccessible. By aligning computational and experimental timescales, this work not only reveals new insights into the mechanics of biomolecular unfolding at realistic velocities but also paves the way for future studies seeking to reconcile theory and experiment in single-molecule force spectroscopy.

Building on these achievements, we continue to push the boundaries of computational power and methodological innovation—developing novel simulation strategies that more closely mimic experimental conditions and yield results directly testable in the laboratory. These techniques are now being applied to a wide range of systems, including dynein motor function, microtubule-associated protein binding, the mechanical properties of PR65, HLA class I peptide binding dynamics, and nanobody conformational changes, offering an unprecedented glimpse into the intricate molecular mechanics that underpin biological function.

Methodological Innovation for Enhanced MD Sampling

Our PI, Dr. Gur, together with his postdoc PI, Dr. Ivet Bahar, developed collective MD (coMD), a hybrid method [4,5] that integrates elastic network models (ENMs) with full-atomic MD simulations through a Monte Carlo–based Metropolis algorithm. This method was recognized as “New & Notable” [6] by the Biophysical Journal and incorporated into widely used tools such as ProDy [7,8] and VMD. By leveraging global, low-resolution “collective modes” (as captured by ENMs), coMD significantly enhances conformational sampling while preserving the full atomic detail of interactions and energetics.

Beyond coMD, Dr. Gur and his lab have advanced simulation methodologies by creatively combining established techniques—such as steered MD, conventional MD, and collective variable (Colvars) simulations— to explore molecular transitions under previously untested simulation conditions. These methodologies/protocols enable more efficient sampling of complex energy landscapes and uncovering otherwise rarely sampled intermediate conformations. First unbiased MD simulations are initiated from experimentally known structural states at both “ends” of a transition. On-the-fly updated collective variables—where a collective variable is a user-defined parameter (e.g., distance, angle, or principal component) that captures essential degrees of freedom relevant to a particular transition—are then applied to guide the rigid-body movement of domains, while allowing the rest of the structure to optimize. These processes generate intermediate conformations, from which a swarm of unbiased long MD simulations is launched to explore the remaining conformational space and to probe transition kinetics.

These in-house developed protocols/methods of advanced MD techniques have been successfully used to model conformational transitions that cannot be sampled with standard MD simulations. For example, we modeled the SARS-CoV-2 spike glycoprotein transitioning between its closed and open states—an essential movement for viral infection—which provided the first fully atomic model of spike protein activation during the pandemic (Figure 2) [9].  Additionally, we modeled the dynein priming stroke, achieving the first all-atom, detailed dynamic simulation of a conformational transition within the dynein mechanochemical cycle [10] (Figure 3).

Building on this foundation, our current efforts focus on applying these in-house–developed protocols to various protein systems and integrating ML/AI methods into them, with the ultimate goal of accelerating discoveries in computational structural biology.

Novel Approaches for Constructing Free Energy Surfaces from MD Simulations

We have developed a practical, straightforward methodology to estimate free energy surfaces by integrating several well-established techniques. In this approach, we perform extensive all-atom MD simulations, and then apply principal component analysis (PCA) to the resulting conformational ensembles. By projecting the configurations onto principal components (PCs) and deriving the corresponding PC-projected distributions, we can employ statistical mechanics principles to calculate free energy surfaces.

Our initial demonstration of this method used microsecond (µs)-length MD simulations of the archaeal aspartate transporter and millisecond (ms)-length simulations of the bovine pancreatic trypsin inhibitor. We later refined the workflow by first determining a conformational transition pathway between different protein states using our hybrid coMD methodology, followed by conventional MD simulations. This enhanced strategy enabled us to map the free energy landscape of the leucine transporter [11], identify its minimum free energy pathway during the functional transition, and reveal previously unresolved intermediates.

Subsequently, our lab expanded this approach by generating intermediate conformations along specific collective variables, thus making our methodology applicable to a broad range of systems with diverse functional motions—examples include the SARS-CoV-2 spike protein [9] (Figure 2) and dynein motor protein [10](Figure 3). This advancement further underscores the versatility of our technique for MD simulations and protein research.

Currently, we use PCA—an unsupervised machine learning method—to project and analyze conformational data. We are working on extending our free energy surface generation strategies to include additional AI and machine learning techniques. By doing so, we aim to continue enhancing both the accuracy and the efficiency of our free energy surface calculations across a wide range of protein systems.