• Login
  • Register

Work for a Member company and need a Member Portal account? Register here with your company email address.

Project

EquiJump: AI-Accelerated Protein Molecular Dynamics Simulation

Copyright

Allan dos Santos Costa

Allan dos Santos Costa

Proteins are the workhorses of the cell, powering the biological processes that make life possible. Their precise choreographies catalyze reactions, transport molecules, and build cellular structures, and modeling these intricate dances is crucial to understanding and manipulating biological systems. 

While Molecular Dynamics (MD) simulation offers a powerful tool to study these movements, its computational demands and latency have limited its widespread use.  Instead, learning from trajectory data, generative AI has emerged as a powerful surrogate for overcoming these limitations and providing faster insights into protein dynamics.

In collaboration with NVIDIA, the MIT Atomic Architects and SISSA, we  recently introduced a new atomistic deep learning model –  EquiJump – for learning and accelerating protein dynamics simulation through large jumps in simulation time. EquiJump leverages Euclidean equivariant neural networks within the generative framework of Stochastic Interpolants for directly transporting 3D all-atom proteins between simulation time steps. 

Copyright

Allan dos Santos Costa

EquiJump efficiently operates on all-atom protein data by treating it as a Tensor Cloud,  or a gas of 3D geometric features.  We specialize our architecture for this multi-typed object, using Two-Sided Stochastic Interpolants for learning feature and coordinate updates for drift and noise that can be integrated from a starting step to generate a next configuration:

Copyright

Allan dos Santos Costa

We deploy EquiJump on learning a transferable model for the dynamics of the 12 fast-folding proteins, finding that the learned shared simulator can stabilize their native basins and accurately reproduce their long-term kinetics while performing large jumps in time. 

Copyright

Allan dos Santos Costa

To evaluate the performance of our model, we consider its long-term dynamical behavior. For that, we useTime-lagged Independent Component Analysis (TICA) and Markov State Modeling to estimate the free energy of structural observables. Our comparisons show that EquiJump can reproduce the long-term dynamics of the proteins, covering their phase spaces, identifying the main basins and matching important relative transition probabilities. 

Copyright

Allan dos Santos Costa

Our results suggest that EquiJump offers a promising avenue for learning accelerated protein dynamics simulators, paving the way for deeper insights into our structural understanding of biology.  As we continue to refine and expand upon this work, we anticipate exciting new frontiers in the realm of biomolecular simulations.

Read more about EquiJump in our preprint paper and stay tuned for updates and future projects!