Comment by edmcnulty101
2 years ago
How bad is our understanding of force fields?
It seems like that's the basic principle to understand.
2 years ago
How bad is our understanding of force fields?
It seems like that's the basic principle to understand.
I think many people would say that in principle, you could make a QM force field with an accurate enough basis function that an infinitely long simulation would recapitulate the energy landscape of a protein, and that information could be used to predict the kinetically accessible structures the protein adopts.
In practice, the force fields are well understood but to be computationally efficient, they have to approximate just about everything. Examples: since number of inter-atom distance pairs goes up with N**2 atoms, you need to have tricks to avoid that and instead scale around n log n or even n if you can do it. When I started, we just neglected atoms more than 9 angstrom apart, but for highly charged molecules like DNA, that leads to errors in the predicted structure. Next, typically the force fields avoid simulating polarizability (the ability of an atom's electron cloud to be drawn towards another atom with opposite charge), also because expensive. They use simplified spring models (lterally hooke's equation) for bond lengths, bond angles. The torsions (the angle formed by 4 atoms in a row) haev a simplified form. The interatomic relationships are not handled in a principled way, instead treating atoms as mushy spheres....
After having made major contributions in this area, I don't think that improvements to force fields are going to be the most effective investment in time and energy. There are other bits of data that can get us to accurate structures with less work.
That's interesting. Didn't realize that. It sounds like we're just working around slower computation speed.
In an fantasy world if we had infinite computation speed/space we'd be able to just model the force field and predict from there.
Yes, that's a fantasy world. I explored this using the Exacycle system at Google and we did actually do a couple things that nobody else could have at the time, but even that extraordinary amount of computing power really is tiny. The problem is the "force field" isn't just the enthalpic contributions I listed above, but also depends intimately on much more subtle entropic details- things like the cost of rearranging water into a more ordered structure have to be paid for. Estimating those is very expensive- far worse than just enumerating over large numbers of proteins "in vacuo", and probably cannot be surmounted, unless quantum computing somehow becomes much better.
Instead, after spending an ordinate amount of Google's revenue on extra energy, I recommended that Google instead apply machine learning to protein structure prediction and just do a better job of extracting useful structural information (note: this was around the time CNNs were in vogue, and methods like Transformers didn't exist yet) from the two big databases (all known proteins/their superfamily alignments, and the PDB).
Note that this conclusion was a really hard one for me since I had dedicated my entire scientific career up to that point in attempting to implement that fantasy world (or a coarse approximation of it), and my attempts at having people develop better force fields (ones that didn't require as much CPU time) using ML weren't successful. What DeepMind did was, in some sense, the most parsimonious incremental step possible to demonstrate their supremacy, which is far more efficient. Also, once you have a trained model, inference is nearly free compared to MD simulations!
1 reply →