BPS2026 – Steering structure predictions of mechanical proteins using twisted diffusion sampling and arbitrary Bayesian likelihoods

Deep learning approaches such as those used by AlphaFold have revolutionized protein structure prediction. These tools are trained using experimental structures from the protein data bank (PDB) and accordingly recapitulate the conformations available in that database. There is great interest in predicting conformations that may be underrepresented in the PDB and that are functionally relevant, such as stretched states of mechanical proteins and open states of mechanosensitive channels. Here we exploit the fact that many modern structure prediction tools, such as AlphaFold3 and the Boltz-2 MIT open-source reproduction of it, use generative artificial intelligence diffusion models to predict a protein’s atomic coordinates from evolutionary sequence information. We reframe the search for alternative and stretched molecular conformations as that of sampling from a conditional diffusion distribution that is formed using an arbitrary Bayesian likelihood specified by the user. Inside of Boltz-2, we implement a twisted diffusion sampler, which is a hybrid of classifier guidance and sequential Monte Carlo strategies, to efficiently sample this conditioned distribution. We demonstrate the utility of this approach by implementing a generative artificial intelligence diffusion analog of steered molecular dynamics simulations applied to benchmark systems such as the muscle protein titin, the inner-ear tip link protein protocadherin-15, and the bacterial channel of large conductance MscL. We are able to reproduce stretched and unfolded states of titin and protocadherin-15, as well as open states of MscL consistent with experimental results. We expect that steered structure predictions will help sample underrepresented and non-equilibrium conformations for a variety of macromolecular systems.