Abstract: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. 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. Despite the lack of molecular dynamics (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. Our study demonstrated that DeepConformer learned energy landscape can be used to efficiently explore protein conformational distribution and dynamics.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the key problems in protein structure prediction and dynamic simulation. Specifically: 1. **The gap between static structure and dynamic behavior**: - Although existing protein structure prediction methods (such as AlphaFold2) have made revolutionary progress in predicting a single static structure, they mainly focus on predicting a single experimental - level structure from the amino acid sequence. However, the function of a protein depends not only on its static structure but also on its conformational distribution and dynamic behavior in solution. - The paper points out that understanding the dynamic characteristics of proteins is crucial for revealing their biological functions. Therefore, a method that can effectively sample the conformational distribution of proteins and simulate their dynamic behavior is needed. 2. **Learning energy landscapes and conformational dynamics from experimental structures**: - The authors assume that the energy landscapes and conformational dynamics of proteins can be learned from the experimental structures in the PDB (Protein Data Bank) and co - evolution data. This means that without relying on computationally intensive molecular dynamics (MD) simulation data, the conformational flexibility and dynamic characteristics of proteins can also be captured. - To this end, they developed a diffusion - generation model named DeepConformer, which can sample the conformational distribution of proteins from a given amino acid sequence. 3. **Improving the sampling ability for complex conformational changes**: - In order to enhance the model's ability to learn the protein energy landscape, DeepConformer introduces a variety of techniques, such as correlating different structures, masking a large proportion (50 - 70%) of amino acid positions on a large scale, and using multiple - sequence alignment (MSA) clustering methods. - These techniques enable DeepConformer not only to generate conformations close to known structures but also to explore the transition paths of proteins between different energy basins, thereby better simulating complex conformational changes. ### Summary The core problem of the paper is: how to use deep - learning methods, especially diffusion - generation models, to learn the energy landscapes and conformational dynamics of proteins from experimental structures and co - evolution data, so as to make up for the deficiencies of existing static - structure prediction methods in dynamic - behavior simulation. Through this method, researchers hope to more comprehensively understand the dynamic characteristics of proteins and their biological functions.

Deep learning of protein energy landscape and conformational dynamics from experimental structures in PDB

Exploring Protein Conformational Changes Using a Large‐Scale Biophysical Sampling Augmented Deep Learning Strategy

ExEnDiff: An Experiment-guided Diffusion model for protein conformational Ensemble generation

Machine Learning Generation of Dynamic Protein Conformational Ensembles

Mapping transiently formed and sparsely populated conformations on a complex energy landscape

Sampling Conformational Ensembles of Highly Dynamic Proteins via Generative Deep Learning

Exploring Multi-Dimensional Coordinate-Dependent Diffusion Dynamics on the Energy Landscape of Protein Conformation Change

Direct generation of protein conformational ensembles via machine learning

EigenFold: Generative Protein Structure Prediction with Diffusion Models

DynamicBind: predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model

Protein Conformation Generation via Force-Guided SE(3) Diffusion Models

AlphaFolding: 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance

Exploring the conformational ensembles of protein-protein complex with transformer-based generative model

Protein Structure Prediction: Conventional and Deep Learning Perspectives

Dynamic PDB: A New Dataset and a SE(3) Model Extension by Integrating Dynamic Behaviors and Physical Properties in Protein Structures

Conformational landscapes of membrane proteins delineated by enhanced sampling molecular dynamics simulations

Deep learning for protein structure prediction and design—progress and applications

Predicting protein conformational motions using energetic frustration analysis and AlphaFold2

Prediction of Molecular Conformation using Deep Generative Neural Networks

Towards predicting equilibrium distributions for molecular systems with deep learning

Predicting equilibrium distributions for molecular systems with deep learning