北京大学、昌平实验室以及哈佛大学的研究团队提出了一种名为ColabDock的通用框架，用于综合结构预测蛋白-蛋白对接，这一研究成果发表于《Nature Machine Intelligence》杂志。该框架采用深度学习结构预测模型，能够整合不同形式和来源的实验约束，无需进行大规模的再训练或微调。
ColabDock的表现优于使用AlphaFold2作为结构预测模型的HADDOCK和ClusPro，尤其是在复杂结构预测中。它还可以通过模拟界面扫描限制来帮助抗体-抗原界面预测。这项研究对药物研发、抗体设计等应用领域具有重要意义，因为蛋白质复合物结构预测在其中的作用至关重要。尽管深度模型在蛋白质结构预测方面发展迅速，但大多数模型都是以自由对接的方式进行预测，这可能会导致实验约束与预测结构不一致。ColabDock的提出，旨在弥合实验与计算之间的差距，通过梯度反向传播，有效地整合了实验约束的先验和数据驱动的蛋白质结构预测模型的能量景观，自动搜索满足两者的构象，同时容忍约束中的冲突或模糊性。
ColabDock框架包含两个阶段：生成阶段和预测阶段。在生成阶段，ColabDock采用了基于AlphaFold2的蛋白质设计框架ColabDesign。在预测阶段，根据生成的复合物结构和给定的模板预测结构。研究人员用合成数据集和几种类型的实验约束上测试ColabDock，包括NMR化学位移扰动(CSP)、共价标记(CL)和模拟深度突变扫描(DMS)。在基准测试集上的测试结果表明ColabDock取得了令人满意的性能，并在提供更多限制的情况下收敛到更少的构象，表明有效应用了附加信息。与HADDOCK和ClusPro相比，ColabDock在实验数据集上的表现更为突出，无论提供的约束数量和质量如何，ColabDock的性能仍然优于HADDOCK和ClusPro。
这一研究成果为理解生物机制提供了重要的结构信息，同时也为人工智能在药物研发等领域的应用提供了新的工具和方法。随着ColabDock框架的不断发展和完善，其在生物医药领域的影响力有望进一步扩大。
英语如下：
Title: “Beijing University Team Breakthrough: AI Framework Accurately Predicts Protein-Protein Docking Structures”
Keywords: BNU AI, Protein Docking, Experimental Prediction
Content: A research team from Peking University, the Changping Laboratory, and Harvard University has proposed a universal framework named ColabDock for comprehensive structural prediction of protein-protein docking. This research was published in the journal Nature Machine Intelligence. The framework utilizes a deep learning structural prediction model to integrate various forms and sources of experimental constraints without the need for extensive retraining or fine-tuning.
ColabDock outperforms HADDOCK and ClusPro, which use AlphaFold2 for structural prediction, especially in complex structure prediction. It can also assist in predicting antibody-antigen interface through simulated interface scanning limitations. This study is of great significance for applications such as drug development and antibody design, as protein complex structure prediction plays a crucial role in these areas. Despite the rapid development of deep models in protein structure prediction, most models predict in a free-docking manner, which may lead to inconsistencies between experimental constraints and predicted structures. The introduction of ColabDock aims to bridge the gap between experiment and computation, effectively integrating experimental constraints with the energy landscapes of data-driven protein structure prediction models through gradient-based backpropagation, and automatically searching for conformations that satisfy both, while tolerating conflicts or ambiguities in the constraints.
The ColabDock framework consists of two stages: a generation stage and a prediction stage. During the generation stage, ColabDock adopts the protein design framework ColabDesign based on AlphaFold2. In the prediction stage, the framework predicts structures of the complex based on the generated compound structure and given templates. Researchers tested ColabDock on synthetic datasets and several types of experimental constraints, including Nuclear Magnetic Resonance (NMR) chemical shift perturbations (CSP), covalent labeling (CL), and simulated deep mutational scanning (DMS). Testing results on a benchmark test set indicated that ColabDock achieved satisfactory performance and converged to fewer conformations with more constraints, indicating effective application of additional information. Compared to HADDOCK and ClusPro, ColabDock’s performance is more outstanding on experimental datasets, maintaining superior performance regardless of the number and quality of constraints provided.
This research provides important structural information for understanding biological mechanisms and offers new tools and methods for the application of artificial intelligence in fields such as drug development. With the continuous development and improvement of the ColabDock framework, its influence in the biomedical field is expected to expand further.
【来源】https://www.jiqizhixin.com/articles/2024-08-07-6
Views: 2