Researchers at Auburn University and Colorado State University, supported by the National Science Foundation, have advanced our understanding of mechanostability in proteins by integrating AI models into computational biophysics. In our earlier news report, we highlighted how their work sheds light on the ultrafast activation of catch bonds and challenges classical assumptions about protein unfolding and stability. Now, we sat down with members of the research team to explore how they combined large-scale molecular simulations with machine learning, the unexpected diversity of binding modes they uncovered, and what their findings could mean for future biomedical and biotechnological applications.
You can see the full research paper here:
C. R. Melo, M., & Bernardi, R. C. (2025). AI Uncovers the Rapid Activation of Catch-Bonds under Force. Journal of Chemical Theory and Computation. https://doi.org/10.1021/acs.jctc.5c01181
You can also check out an additional published material below:
Melo, M. C. R., Gomes, D. E. B., & Bernardi, R. C. (2023). Molecular Origins of Force-Dependent Protein Complex Stabilization during Bacterial Infections. Journal of the American Chemical Society, 145(1), 70–77. https://doi.org/10.1021/jacs.2c07674
Follow some of the authors here:
The following interview is presented unedited to preserve the team’s original insights, giving you a direct view into how AI-driven simulations are redefining our understanding of protein mechanics and opening new avenues for therapeutic innovation.
Acknowledgment from Marcelo Melo, Rafael Bernardi and team: The authors would like to acknowledge the support from Auburn University and Colorado State University, as well as funding from the National Science Foundation under Grant MCB-2143787.
Could you describe how AI models (e.g., machine learning or deep learning) were integrated into your computational biophysics workflow to predict or uncover rapid protein bond activation events?
We used AI models primarily to test a hypothesis related to how soon a catch bond is activated. The basic idea is simple: If the protein “decided” on a level of mechanical stability early in the simulation, we would be able to predict their force required for rupture b y looking at the beginning of the simulations. By running hundreds of independent tests and using them as training/testing data, we could train AI models to predict ruptures forces, confirming the hypothesis that catch bonds are activated very quickly in this molecular system.
Among the different bond types and protein environments you probed, which specific bonds exhibited the highest sensitivity to environmental factors (e.g., pH, temperature), and what did the AI reveal about their dynamic behaviour?
The interface between the two proteins we tested revealed itself to be highly dynamic, and throughout the hundreds of simulations we performed, we could observe a variety of binding modes, which included hydrogen-bonds, polar interactions, and hydrophobic contacts. However, our tests did not reveal any clear pattern in bond types.
How did you validate the AI-predicted activation pathways? Did you rely on molecular dynamics simulations with enhanced sampling techniques, quantum mechanical/molecular mechanical (QM/MM) calculations, or experimental collaboration (e.g., spectroscopy)?
We chose a molecular system that had been thoroughly explored through experiments, including single-molecule-force-spectroscopy, so we could count on ample experimental data on binding constants and mechanical stability.
What role did conformational flexibility or local structural motifs (e.g., loops, active sites, hydrogen bonds) play in modulating bond activation, according to your AI-derived insights?
Our AI models showed that two binding interfaces were important for the mechanical stability of the complex, on between the two proteins (cohesin and dockerin) and another between subdomains of the dockerin protein (the X-module and the binding domain). Only by including both binding interfaces in our training data could the AI models reliably predict rupture forces. This highlights how the internal dynamics of the dockerin protein (not just the interface between proteins) was essential for the activation of this catch bond.
How do your findings refine the current understanding of enzymatic catalysis or protein stability? For instance, do they suggest new models for the coupling between protein motion and chemical reactivity?
This complex does not have catalytic activity, but did present an open question regarding unfolding being necessary for the strong mechanical resiliency of the complex. In our study, we saw the activation of the catch bond occurring without unfolding of the protein, which suggests a new model for the distribution of forces throughout the complex, making it mechanostable even without partial unfolding of the dockerin protein.
Were there surprises in the activation mechanisms you uncovered; such as previously overlooked pathways or ultrafast events, that challenge classical kinetics models in biophysical chemistry?
One of the most surprising results of our study was the remarkable flexibility and diversity of binding modes between cohesin and dockerin, even though the pair of proteins forms one of the strongest protein complexes ever recorded. This challenges our assumptions regarding a conserved binding mode being required for mechanically strong molecular complexes.
Looking ahead, how might you extend this AI-guided approach to study ligand binding, allosteric regulation, or disease-associated mutations, and what experimental collaborations would help translate your computational findings into practical applications?
We are currently working with experimental collaborators to pursue a new understanding of mechanostability in biomolecules. In particular, we are looking into how force redistribution can change complex stability, and how protein dynamics can reveal different binding modes. This will have a direct impact in biotechnological and biomedical applications where mechanostable binding (such as antibody-directed drug delivery) and force-responsive dynamics (such as in force activated catalysis) have the potential to create new technologies and treatments. Moreover, predicting mechanostability from simulations can help us understand how proteins evolved to become so mechanically resilient.
Hassan graduated with a Master’s degree in Chemical Engineering from the University of Chester (UK). He currently works as a design engineering consultant for one of the largest engineering firms in the world along with being an associate member of the Institute of Chemical Engineers (IChemE).
