Physics-informed deep learning approach for fixing and sampling protein 3D structures

This is the authors' abstract. We don't add key points for this paper.

AlphaFold2 (AF2), the first deep learning-based protein structure prediction model to achieve near-experimental accuracy, has revolutionized the field of structural biology. With its ability to predict protein structures at unprecedented levels of accuracy, AF2 has become an essential tool in structural biology. In parallel, the deep learning methodologies derived from AF2 have made significant advancements in a range of downstream applications, such as efficient sampling of protein conformational ensembles, protein sequence design, and ligand generation. These applications are underpinned by an accurate description of the protein-free energy landscape, which is critical for modeling protein behavior in various biological contexts. Despite these advancements, challenges remain, particularly in the area of multi-conformation sampling. Traditional methods such as molecular dynamics (MD) and Monte Carlo (MC) simulations, which are based on Boltzmann statistics, can provide detailed insights into conformational distributions. However, the application of these methods is hindered by their high computational cost and limited efficiency, especially for large or complex systems. In contrast, existing generative deep learning models can rapidly generate protein conformations, but often struggle to incorporate biophysical constraints, and are limited in the power of predicting the accurate energy minima, avoiding steric clashes, or understanding physics such as electrostatic interactions. This lack of control over the physical space makes it difficult to generate conformations that are both physically realistic and biologically relevant. Furthermore, the robustness of the initial protein structure is crucial for downstream applications, including protein-protein interaction predictions, molecular docking, and virtual screening, as well as reliable and efficient MD simulations. However, conventional structure repair tools often fail to accurately model missing regions, such as loops or partially resolved regions, which may lead to unrealistic geometries or energetically unfavorable conformations. To address these limitations, we present FoldCopilot, an innovative suite based on AF2. FoldCopilot uses both target sequences and multiple sequence alignments (MSAs) as input prompts for the Evoformer module, leveraging a homology-based sequence perturbation mechanism that generates diverse local conformations without the need for retraining. This method allows for the efficient generation of protein conformations and at the same time maintains high fidelity with known structural templates and evolutionary constraints. In addition, FoldCopilot integrates structural templates and initial reference frames into the iterative refinement process, enabling the directed evolution and control over generated protein structures. This iterative process ensures that the generated conformations not only adhere to biophysical principles, such as realistic energy states and steric compatibility but also meet specific functional requirements dictated by the target application. Our experimental results demonstrate that FoldCopilot excels in key tasks such as protein structure repair and localized multi-conformation generation. The method significantly outperforms traditional structure repair techniques, yielding more accurate and biologically relevant models for regions of missing data. Additionally, FoldCopilot shows a high capability to generate diverse conformations, making it an ideal tool for protein engineering, protein-ligand interaction predictions, and other structure-based drug design applications. In summary, FoldCopilot bridges the gap between high-fidelity conformational sampling and the ability to control conformational space, offering a powerful approach to facilitate protein-molecule interaction predictions, structural modeling, and molecular dynamics simulations.