专访西湖大学卢培龙：AI蛋白质设计目前还无需严格监管，否则可能减缓科学进步

Core Insights - The article discusses the transformative impact of machine learning-based AI tools on protein structure research, highlighting advancements in computational tools for predicting protein structures and properties, and their applications in protein design [1][2][3]. Group 1: Advances in Protein Modeling - Machine learning tools are addressing challenges in understanding macromolecular dynamics and functions, moving beyond single structure limitations [2]. - AlphaFold has revolutionized protein structure biology, making it a common perspective in experimental design [2]. - Recent advancements, particularly AlphaFold3 and RoseTTAFold All-Atom, have significantly improved the accuracy and scope of protein structure and interaction predictions [3]. Group 2: Challenges in Predicting Complex Structures - While tools like AlphaFold-Multimer can accurately predict many tightly interacting complexes, challenges remain in simulating large, dynamic, or transient complexes [4]. - Membrane proteins and intrinsically disordered proteins present significant challenges due to the lack of high-resolution experimental data [4][5]. - The main obstacles to improving predictions for these complex structures include the scarcity of high-resolution experimental data and the need for new deep learning methods [4][5]. Group 3: Dynamic Structures and Environmental Conditions - Current tools do not predict folding pathways and have limitations in considering different solution conditions or temperatures [6][9]. - Integrating multiple data sources into machine learning models is crucial for describing dynamic structures and predicting conformational changes in response to environmental variations [8][10]. Group 4: Tools for Designing Functional Proteins - Current tools struggle to capture the dynamic characteristics of protein functions, which depend on various interactions and require quantitative predictions of conformational changes [10][11]. - Progress has been made in designing tools for properties beyond structure, but challenges remain in accurately predicting binding affinities and controlling dynamic properties [11][12]. Group 5: Integration of Machine Learning with Other Methods - Significant progress has been made in integrating machine learning with molecular dynamics and other computational methods to enhance protein design [17][18]. - AI 2 BMD exemplifies a hybrid system that combines machine learning with quantum mechanics to simulate large biomolecules with high accuracy [18]. Group 6: Future Directions and Data Needs - Expanding datasets to include functional and biophysical measurements is essential for improving model predictions [15][16]. - The development of databases for dynamic and polymorphic conformations is critical for simulating proteins that rely on structural dynamics for functionality [15][16]. Group 7: Responsible Use of AI in Protein Design - Concerns about the responsible use of AI in protein design include the potential for creating harmful proteins and the energy consumption of AI models [19][20]. - Continuous improvement of detection algorithms and collaboration between academia and industry are necessary to ensure safety while promoting scientific advancement [20][21].