Periodic Reporting for period 4 - ScaleCell (Scalable Kinetic Models: From Molecular Dynamics to Cellular Signaling)
Reporting period: 2022-11-01 to 2024-04-30
Proteins can dynamically switch between metastable structures and associate into signaling machineries that give rise to cellular function. Molecular dynamics (MD) simulations can simultaneously probe structure and dynamics of such processes at atomistic resolution. Recently, breakthroughs have been achieved in the long-standing problem to sample rare transition events in unbiased MD. In the past, we have pioneered the development and application of Markov state models (MSMs), which, combined with MD simulations on graphical processing units, make millisecond-timescale kinetics broadly accessible. With multi-ensemble techniques, timescales of seconds and beyond can be reached. Recently, we have demonstrated protein-protein association and dissociation with atomistic MD. Long-timescale simulation of small to medium-sized protein systems is thus now possible in atomistic MD.
However, these methods have fundamental scaling limitations that prevents long-timescale simulation to be employed in the modeling of large protein systems and whole cells. To address these limitations, this ERC project has set out to develop multiscale models for coarse-grained representations of proteins and upscaling methods from the molecular scale to reaction-diffusion simulations of cellular signal transduction.
However, these methods have fundamental scaling limitations that prevents long-timescale simulation to be employed in the modeling of large protein systems and whole cells. To address these limitations, this ERC project has set out to develop multiscale models for coarse-grained representations of proteins and upscaling methods from the molecular scale to reaction-diffusion simulations of cellular signal transduction.
The following key results were achieved on all the planned work packages:
1) Development of a kinetic modeling framework that scales to large molecular systems: We continued developing the Markov state modeling (MSM) framework and extended to Markov random fields which are fundamentally able to describe rare-even transitions / conformational changes in large molecular complexes by means of a coupled field of transitions in individual domains.
2) Development of an upscaling scheme that predicts kinetics of multi-body molecular interactions based on a manageable number of MD simulations of small subsystems: We learned that a natural description of the emerging macroscopic thermodynamics and kinetics in large-scale molecular systems are effective, coarse-grained forcefields. These had previously been difficult to model, but with the deep learning revolution we were in a position to machine-learn them from atomistic simulation data. In collaboration with the Clementi group we formulated a bottom-up machine-learning approach (CGnet) and showed that graph neural networks can effectively learn the effective interaction potential between the particles of a coarse-grained representation of proteins. We further optimized the methodology by automatically choosing a statistically optimal coarse-graining scheme and developing estimators that combing the saved atomistic forces and the sampled distribution for better results. These methods have already attracted significant visibility, citations and follow-up work.
3) Development of a multi-scale scheme that couples molecular kinetics models and particle-based reaction-diffusion dynamics towards cell-scale simulations with molecular detail: We have developed a rigorous mathematical coupling framework between molecular kinetics, as described by MSMs and reaction-diffusion dynamics.
Generally usable software was developed to implement the new technologies and provided open-source to the scientific community, mainly DeepTime (https://deeptime-ml.github.io) and ReaDDy (https://readdy.github.io/).
1) Development of a kinetic modeling framework that scales to large molecular systems: We continued developing the Markov state modeling (MSM) framework and extended to Markov random fields which are fundamentally able to describe rare-even transitions / conformational changes in large molecular complexes by means of a coupled field of transitions in individual domains.
2) Development of an upscaling scheme that predicts kinetics of multi-body molecular interactions based on a manageable number of MD simulations of small subsystems: We learned that a natural description of the emerging macroscopic thermodynamics and kinetics in large-scale molecular systems are effective, coarse-grained forcefields. These had previously been difficult to model, but with the deep learning revolution we were in a position to machine-learn them from atomistic simulation data. In collaboration with the Clementi group we formulated a bottom-up machine-learning approach (CGnet) and showed that graph neural networks can effectively learn the effective interaction potential between the particles of a coarse-grained representation of proteins. We further optimized the methodology by automatically choosing a statistically optimal coarse-graining scheme and developing estimators that combing the saved atomistic forces and the sampled distribution for better results. These methods have already attracted significant visibility, citations and follow-up work.
3) Development of a multi-scale scheme that couples molecular kinetics models and particle-based reaction-diffusion dynamics towards cell-scale simulations with molecular detail: We have developed a rigorous mathematical coupling framework between molecular kinetics, as described by MSMs and reaction-diffusion dynamics.
Generally usable software was developed to implement the new technologies and provided open-source to the scientific community, mainly DeepTime (https://deeptime-ml.github.io) and ReaDDy (https://readdy.github.io/).
Several important and unexpected developments were enabled throughout the ERC project by the deep learning revolution.
1) The most important development are "Boltzmann Generators" which was published 2019 in Science and had received big attention in the molecular simulation community and also the machine learning community. Boltzmann Generators are an approach to sample from the molecular equilibrium distribution with generative deep neural networks without suffering from the sampling problem. We and others have improved and optimized the technology over the years and its most current incarnation are the widely used diffusion models for generating molecular structures. Further research will be needed to mature this line of research into a system that solves the molecular sampling problem in general while keeping very high accuracy, such that molecular dynamics (MD) simulations and also experimental biochemical assays can be effectively replaced by such a deep learning emulator.
2) Several other deep learning developments were made in the scope of the project which go beyond our original plans, including a deep learning approach for solving the electronic structure problem in quantum Chemistry (DeepQMC).
3) Throughout the Corona pandemic we had started a close collaboration with researchers at the National Institutes of Health (Matt Hall lab) and the German Primate Research Center in Göttingen (Stephan Pöhlmann lab). In this collaboration we had used our technologies to simulate the human TMPRSS2 receptor which is important in the Corona infection pathway, as well as in other virus diseases, and developed highly effective (nanomolar) inhibitors that were tested in biochemical and cellular assays by our collaborators.
1) The most important development are "Boltzmann Generators" which was published 2019 in Science and had received big attention in the molecular simulation community and also the machine learning community. Boltzmann Generators are an approach to sample from the molecular equilibrium distribution with generative deep neural networks without suffering from the sampling problem. We and others have improved and optimized the technology over the years and its most current incarnation are the widely used diffusion models for generating molecular structures. Further research will be needed to mature this line of research into a system that solves the molecular sampling problem in general while keeping very high accuracy, such that molecular dynamics (MD) simulations and also experimental biochemical assays can be effectively replaced by such a deep learning emulator.
2) Several other deep learning developments were made in the scope of the project which go beyond our original plans, including a deep learning approach for solving the electronic structure problem in quantum Chemistry (DeepQMC).
3) Throughout the Corona pandemic we had started a close collaboration with researchers at the National Institutes of Health (Matt Hall lab) and the German Primate Research Center in Göttingen (Stephan Pöhlmann lab). In this collaboration we had used our technologies to simulate the human TMPRSS2 receptor which is important in the Corona infection pathway, as well as in other virus diseases, and developed highly effective (nanomolar) inhibitors that were tested in biochemical and cellular assays by our collaborators.