Collaborative Research: Frameworks: Interoperable High-Performance Classical, Machine Learning and Quantum Free Energy Methods in AMBER

点击翻译按钮获取中文摘要

With support from the Office of Advanced Infrastructure and the Division of Chemistry at NSF, Professor Merz and his group will work on molecular simulation cyberinfrastructure. Molecular simulations have become an invaluable tool for research and technology development in chemical, pharmaceutical, and materials sciences. With the availability of specialized hardware such as graphics processing units (GPUs), molecular dynamics simulations using classical or molecular mechanical force fields have reached the spatial and temporal scales needed to address important real-world problems in the chemical and biological sciences. Free energy simulations are a particularly important and challenging class of molecular simulations that are critical to gain a predictive understanding of chemical processes. For example, free energy methods can predict the barrier height and rates for chemical reactions, whether a reaction will occur, or how tightly a drug binds to a target. These predictions are extremely valuable for the design of new catalytic agents or drugs. However, the predictive capability of free energy simulations is sensitive to the underlying model that describes the inter-atomic potential energy and forces. Accurate free energy simulations of chemical processes require potential energy models that capture the essential physics and can respond to changes in the chemical environment, but conventional force field models are unsuitable for many processes involving bond breaking and formation as seen, for example, in catalyst design. Consequently, there is great need to extend the scope of free energy methods by enabling the use of a broader range of potential energy models that are more accurate as well as reactive and/or capable of quantum mechanical many-body polarization and charge transfer. The cyberinfrastructure created by this project allows for the routine application of free energy methods, using quantum mechanics, machine learning, reactive and classical potentials to a myriad of important problems that advance the state-of-the art in the biological and chemical sciences. The tools can be applied by a range of scientists to address fundamental problems of national interest, for example, in the design of drugs against zoonotic diseases (e.g., COVID-19), the design of materials with novel functions and in the design of improved batteries. Given the sophistication of the methods employed, education of a diverse pool of chemical, biological and computer scientists to advance this field is essential and is addressed in this project, thereby training the next generation of computational scientists that will form the backbone of the work force of the future. The project develops accurate and efficient free energy software within a powerful new multiscale modeling framework in the AMBER suite of programs for applications in chemistry, biology, and materials science. The multiscale framework enables the design and use of new classes of mixed-method force fields that involve interoperability between several existing and emerging reactive, machine learning and quantum many-body potentials. These potentials have enhanced accuracy, robustness, and predictive capability compared to classical molecular mechanical force fields and enable the study of chemical reactions and catalysis. The cyberinfrastructure supports innovative multi-layered hybrid potentials that can be customized to meet the needs of complex applications in biotechnology development, enzyme design and drug discovery. A robust endpoint "book-ending" approach that leverages the GPU-accelerated capability of the AMBER molecular dynamics engine is used to reach these goals. Specifically, the open-source high-performance software for free energy simulations is designed for multi-layered hybrid potentials using combinations of linear-scaling many-body quantum mechanical methods via the GPU-accelerated QUICK package, scalable reactive ReaxFF force fields via the PuReMD package, as well as the recently developed DeepMD-SE, ANAKIN-ME (ANI) and AP-Net families of machine learning potentials. The cyberinfrastructure is built upon the existing high-performance CUDA MD engine in AMBER and extends it to a broad range of GPU-accelerated architectures using industry-standard programming models. Scalability is ensured using innovative parallel algorithms. High impact is achieved by leveraging AMBER's broad user base to expand the scope and success of FE applications. In this way, the project leverages existing recognized capabilities and actively engages a diverse team of collaborators and the broader molecular simulations community. The cyberinfrastructure delivered by the project enables a wide range of new and enhanced applications for a broad community of users in academia, industry, and national laboratories. These applications include drug discovery, enzyme catalysis, and biomaterials design. The AMBER suite of programs has a long-standing extensive worldwide userbase, and is widely used on national production cyberinfrastructure. The enhancement of AMBER as an established, proven sustainable, and widely used package will ensure that the software has a broad impact well beyond the end of the project. The project will also train a diverse population of students and researchers in theory, programming, computational chemistry/biology, computer science, scientific writing, and communication.This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Division of Chemistry within the NSF Directorate for Mathematical and Physical Sciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.