Transferable Adaptive Conformation Dynamics

• Applicant: Professor Dr. Frank Noé
• Subject Area: Biophysics
• Term: since 2011
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 200023966

Life processes such as growth and metabolism emerge from the interactions of proteins and other biomolecules. Proteins can store and process information by switching between conformations, i.e., long-lived sets of 3D structures, and can associate and dissociate with other biomolecules, to build up biological structures and process cellular signals. Biology is inherently multi-scalar: atomic details at key interactions are critical for the macroscopic system behavior. In order to control biological processes, in the design of drugs, biomaterials or biotechnological processes, for example, knowledge of the relevant atomic-detail structures, their populations and transition rates is a key advantage.Molecular dynamics (MD) simulation can probe all-atom structure and dynamics simultaneously, and state-of-the-art simulations can make qualitative and quantitative predictions, thus being an important complement to experimental studies. However, one of the key limitations of MD is the so-called sampling problem: transitions between functionally relevant states of biomolecules, such as protein folding, or the binding of a drug molecule to a protein, are rare events. With direct MD simulations these processes may require years to simulate even on a supercomputer. In the past few years, significant advances have been made on addressing the sampling problem. For example, by using adaptive simulation methods developed in the first funding period of the present project, the PI and his collaborators have achieved the first all-atom simulation of the full association and dissociation process of two proteins. Despite this success, the approach required a computational time of 20,000 days on high-end graphics processors, and is thus still unsustainably expensive.For the second funding period, we propose to develop an adaptive simulation method for MD that uses a fundamentally new approach for sampling. Instead of exclusively relying on methods that try to speed up the MD sampling of a given biomolecules, we will attempt to predict the long-lived states, their probabilities and transition rates of proteins that we have not simulated yet. The proposed method combines ideas from machine learning for quantum chemistry, where predictions across chemical space have been successfully made, with our own preliminary work of using machine learning for computing the long-time statistical behavior of proteins. If successful in making predictions across protein sequences, this method can be of extraordinary usefulness for protein design, e.g., in the context of optimizing enzymes or designing antibodies.

• DFG Programme: Research Grants