Title:

Physics aware deep learning for modeling macromolecular interactions

Abstract:

Macromolecular interactions are building blocks of functioning living organisms and 3D structure of such molecular complexes are critically important for drug discovery. These interactions are driven by laws of physics, but the systems are exceedingly complicated to be fully described from first principles. Most recently deep learning made tremendous progress in the field for modeling structure of some of macromolecules such as proteins, however many pharmaceutically important problems remain unsolved. We have recently shown that combination of deep learning with physics base simulation improved the overall results. We propose to go one step further, and design deep learning architectures encoding geometry and physics of the system , which mimics physical simulation, and learns correction parameters as part of the training. We observe improved results as compared to regular simulation.

Short bio

Dima Kozakov is Frey Family Foundation Professor of Applied Mathematics and Computer Science, with appointments at IACS, Laufer Center, and Institute for Drug Discovery. His research focus is in the algorithm development in computational biology, with the focus on computational effectiveness and innovation. Dr. Kozakov's algorithms and software on modeling macromolecular interactions are licensed by largest pharmaceutical software vendor Schrodinger and installed in all major pharmaceutical companies worldwide. His work is supported by NSF Computer and Information Science (CISE) , and Math and Physical Sciences (MPS) directorates, and National Institute for Health. His recent graduate students are competitively recruited by major Pharma companies, as well as big tech companies such as Google, Microsoft.