As Simple As Possible, but Not Simpler: Exploring the Fidelity of Coarse-Grained Protein Models for Simulated Force Spectroscopy

Mechanical unfolding of a single domain of loop-truncated superoxide dismutase protein has been simulated via force spectroscopy techniques with both all-atom (AA) models and several coarse-grained models having different levels of resolution: A Gō model containing all heavy atoms in the protein (HA-Gō), the associative memory, water mediated, structure and energy model (AWSEM) which has 3 interaction sites per amino acid, and a Gō model containing only one interaction site per amino acid at the Cα position (Cα-Gō). To systematically compare results across models, the scales of time, energy, and force had to be suitably renormalized in each model. Surprisingly, the HA-Gō model gives the softest protein, exhibiting much smaller force peaks than all other models after the above renormalization. Clustering to render a structural taxonomy as the protein unfolds showed that the AA, HA-Gō, and Cα-Gō models exhibit a single pathway for early unfolding, which eventually bifurcates repeatedly to multiple branches only after the protein is about half-unfolded. The AWSEM model shows a single dominant unfolding pathway over the whole range of unfolding, in contrast to all other models. TM alignment, clustering analysis, and native contact maps show that the AWSEM pathway has however the most structural similarity to the AA model at high nativeness, but the least structural similarity to the AA model at low nativeness. In comparison to the AA model, the sequence of native contact breakage is best predicted by the HA-Gō model. All models consistently predict a similar unfolding mechanism for early force-induced unfolding events, but diverge in their predictions for late stage unfolding events when the protein is more significantly disordered.Author Summary: Although experimentalists can now unfold single proteins in the lab by pulling them apart and measuring the force and extension, a clear idea of how the protein changes shape and loses structure during this process is currently missing. Molecular dynamics simulations can offer insight as to what is actually happening structurally when you pull a protein apart. However, typical simulations of processes that happen nearly instantaneously in the lab take weeks to perform, when every atom must be accounted for. Researchers have thus resorted to much faster “coarse-grained models”, where the system is simplified by removing select atoms and the remaining interactions rescaled, but the accuracy of such simulations are known to suffer as a result. How accurate or inaccurate are the current coarse-grained models in capturing the unfolding mechanisms of proteins? Our findings upon investigating this question suggest that, while coarse-grained models successfully capture early unfolding events of nearly-folded proteins, they suffer when trying to describe the late stages of unfolding in mostly-disordered proteins. By showing how coarse-grained models may fail to capture the accuracy of their more sophisticated but cumbersome counterparts, we can shed light on how to improve their reliability, increase their speed, and enhance their relevance in capturing biologically-relevant phenomena.