BPS2026 – Identifying protein unfolding transition states via molecular dynamics

The three-dimensional structure of a protein’s transition state provides insight into protein stability, folding, unfolding, misfolding, and degradation. However, successfully identifying these structures continues to be challenging, both in vitro and in silico, due to physical limitations and stochastic variation. To address this, we analyze replicate simulation trajectories and use complementary techniques to identify the particular frame of a molecular dynamics (MD) simulation corresponding to the transition state. Using all-atom, explicit solvent MD simulations at high temperatures, we have identified the transition state structures of 27 proteins whose transition states have been characterized experimentally through φ value analysis. In doing so, we identified the unfolding transition state in our simulations using two different methods. First, we performed a three-dimensional multidimensional scaling of the all-vs-all Cα RMSD matrix and subsequently identified the exit of the native cluster. Additionally, we used ProteinMPNN to predict sequences that would fold into snapshots from the simulation, then AlphaFold to predict the structure from the multiple sequence alignment. The transition state was identified as the last point at which a structure could be predicted within 5 Å of any structure in the simulation. After isolating a transition state, we used two metrics to quantify how similar the protein was to the native structure and compared these measurements with the experimental φ values. The semiquantitative structure index (S-value) accounts for a protein’s backbone conformation (S2°) as well as tertiary packing (S3°), and lDDT is a superposition-free score that evaluates local distance differences between atoms in a model and the reference structure.