Abstract Background Protein structure prediction has reached revolutionary levels of accuracy on single structures, implying biophysical energy function can be learned from known protein structures. However apart from single static structure, conformational distributions and dynamics often control protein biological functions. Alphafold2/3 currently predict static protein structure only. Several machine learning approaches have been developed to train model using conformations generated from molecular dynamics (MD) simulations. Methods In this work, we tested a hypothesis that protein energy landscape and conformational dynamics can be learned from experimental structures in PDB and coevolution data. Towards this goal, we develop DeepConformer, a diffusion generative model for sampling protein conformation distributions from a given amino acid sequence. We combined three approaches to allow deep learning techniques to extract hidden dynamics energy landscape information: expanded sequence-structure mapping, large scale 50% structure masking, and MSA clustering.Results: Despite the lack of MD simulation data in training process, DeepConformer captured conformational flexibility and dynamics (RMSF and covariance matrix correlation) similar to MD simulation and reproduced experimentally observed conformational variations.DeepConformer can generate conformation close to different native structures and locate intermediate pathway conformations, as illustrated in the Fold switch of KaiB protein (Figure 1).In the case of large conformation change of an interferon-inducible DNA-sensor protein IFI16, Deepconformer predicted close and open conformer transition, which is hard to obtain using MD simulation.For intrinsically disordered protein, Deepconformer generated conformation ensembles agree with experimental conformation distributions. Conclusion Our study demonstrated that DeepConformer learned energy landscape can be used to efficiently explore protein conformational distribution and dynamics. DeepConformer-generated structures has similar dynamic properties to that of MD simulation sampled structures and can cover distinct native structures of a single sequence. As DeepConformer achieved this without using protein-specific models or training with data like molecular dynamics trajectories, we expect it to be widely applicable for exploring the conformational dynamics of both natural and designed proteins to understand and optimize their function and regulation. In IVD applications, many biomarkers have flexible and even disordered conformations. Deepconformer can help to understand the biomarkers’ functions and engineering detection methods.
Ma et al. (Wed,) studied this question.