Misconceptions of Time
Getting the molecular dynamics car out of the garage
For those who are not practitioners of dynamical simulation methods, such as molecular dynamics (MD), one of the biggest misconceptions relates to time. Specifically, the mismatch between the timescales that the simulation can reach compared to what is experimentally relevant. Indeed, typical MD simulations are in the nanosecond to microsecond timescale regime. If the desired phenomenon of interest occurs on the second timescale, one would never see it.
This mismatch often leads people to say that these simulations don’t work, whereas they’re often doing exactly what they should be—reporting on the timescales that they purport to cover. In this sense, it’s like saying that one’s car doesn’t work as a transportation device when you only let it run a few seconds and you never leave the garage.
Once one recognizes this challenge, the natural next step is to devise means to defeat it—to get that car out of the garage, so to speak. In this issue of BCR, we’re featuring a story that digs into this problem of time in dynamical simulations. In recent years we’ve seen a revolution in molecular dynamics simulations in this regard—one headed in multiple yet potentially complementary directions. Most of these approaches, at heart, depend on two key assumptions: 1) Typical atomistic simulations have detail that in some sense is not needed, and 2) One can build such models in a systematic, transparent and reproducible manner.
This duality is apparent in modern coarse-graining methods that systematically devise simpler models of molecular interactions by building up from atomistic simulations. For example, instead of representing a protein as a set of atoms, each modeled by its own sphere and interatomic interactions, one may set a single amino acid molecule (composed of several atoms) as the base unit, thus modeling at a much coarser scale. While coarse graining itself is quite old, what sets modern coarse-graining methods apart is the systematic derivation of models from more detailed models. For example, one might run an atomistic simulation for some time in order to derive the parameters for a simpler, coarse-grained model. This aspect of the work marks a shift away from intuition-based modeling toward data-driven, systematic methods, which has many appealing aspects.
Another approach has been to coarse grain not the interactions but the dynamics itself. If one cares about the millisecond (10-3 seconds) timescale, why would we want to run molecular dynamics with femtosecond (10-15 seconds) scale dynamical steps? One example of this approach is the use of Markov State Models which throw out the uninteresting, very fast timescales (femtoseconds to nanosecond) to gain a dramatic efficiency in calculations, especially in terms of parallelization, i.e., using many short trajectories to reproduce very long timescales. Generating 1000 trajectories each at the microsecond timescale and using them to predict the millisecond timescale is considerably more tractable than a single trajectory on the millisecond timescale.
It’s interesting to consider the future of these approaches. As both spatial and temporal coarse-graining methods become even more systematic and statistically driven, one can imagine how they can start to merge to build the best physical model possible. This future gets particularly exciting when one considers how the “big data” approaches that are starting to revolutionize other fields might also impact dynamical simulation. With such a combination, reaching the millisecond timescale—now out of reach for all but a few researchers—could become routine, enabling molecular simulation to achieve goals that are barely conceivable today.
Vijay S. Pande is professor of chemistry, structural biology and computer science at Stanford University.