PROJECT SUMMARY / ABSTRACT OpenMM [http://openmm.org] is the most widely-used open source GPU-accelerated framework for biomolecular modeling and simulation (>1300 citations, >270,000 downloads, >1M deployed instances). Its Python API makes it widely popular as both an application (for modelers) and a library (for developers), while its C/C++/Fortran bindings enable major legacy simulation packages to use OpenMM to provide high performance on modern hardware. OpenMM has been used for probing biological questions that leverage the $14B global investment in structural data from the PDB at multiple scales, from detailed studies of single disease proteins to superfamily-wide modeling studies and large-scale drug development efforts in industry and academia. Originally developed with NIH funding by the Pande lab at Stanford, we aim to fully transition toward a community governance and sustainable development model and extend its capabilities to ensure OpenMM can power the next decade of biomolecular research. To fully exploit the revolution in QM-level accuracy with machine-learning (ML) potentials, we will add plug-in support for ML models augmented by GPU-accelerated kernels, enabling transformative science with QM-level accuracy. To enable high-productivity development of new ML models with training dataset sizes approaching 100 million molecules, we will develop a Python framework to enable OpenMM to be easily used within modern ML frameworks such as TensorFlow and PyTorch. Together with continued optimizations to exploit inexpensive GPUs, these advances will power a transformation within biomolecular modeling and simulation, much as deep learning has transformed computer vision.

PROJECT NARRATIVE Biomolecular modeling and simulation is a key technology for leveraging the $14B global investment in biomolec- ular structure data in the protein databank to understand the basic molecular mechanisms underlying biology and disease and the development of new therapies. In this proposal, we aim to expand the development of OpenMM, a free and open source biomolecular modeling and simulation package that can exploit a wide range of consumer-grade and high-end graphics processing units (GPUs) to enable researchers and applications built on OpenMM to achieve high performance with extreme flexibility. A key aspect of this proposal is to accelerate research in the emerging field of biomolecular machine learning by tightly integrating OpenMM with modern ma- chine learning frameworks, enabling researchers to build, use, and deploy machine learning potentials, collective variables, and integrators to advance the state of biomolecular modeling.