Metamorphic proteins (MPs) can fold into two or more distinct spatial structures. Increasing interest in MPs has spurred the search for computational tools to predict proteins fold-switching potential and model their refolding pathways. Here we address this problem by using the recently reported generative diffusion predictor UFConf, based on the AlphaFold2 network. We have developed a new UFConf-driven algorithm dubbed IMSD (iterative modeling via structural diffusion) to model the MP's path from one conformational state to another. In brief, we begin with the experimental structure of state A, perturb it through the “noising” process, and infer a number of models (replicas) through the reverse diffusion or “denoising” process. From this set of models, we choose the one that is closest to the alternative structure B; then we use it as a starting point to perform another round of noising/denoising and thus generate the next batch of replicas. Repeating this process in an iterative fashion, we have been able to map the entire path from state A to state B for metamorphic proteins GA98, SA1 V90T, and the C-terminal domain of RfaH. The obtained representation of the fold-switching pathways in these MPs is consistent with the dual-funnel energy landscape observed in the previous modeling studies and shows good agreement with the available experimental data. The new UFConfbased IMSD protocol can be viewed as a part of the emerging generation of modeling tools aiming to model protein dynamics by means of deep learning technology.
