| | (General discussion) | Rosemary charges: graph net based or library charges, general direction, infrastructure needs DM: Pros and cons of using the new graphnet charges which might be unproven conventional library charges, not a burden on infrastructure Right now we have major major issues with conformation dependence of charges and backend-dependence of charges (RDKit and OpenEye won’t give same charges and… we might have to fix that if we keep this). GraphNet-based charges should fix that.
JW: If we get accurate graphnet charges implemented then users need not wait for AM1 charges to be generated. Big cons are pytorch or some ML library becomes a hard dependency. “gold medal”: fully working graphnet “silver medal”: library charges with automatic cap+charge for unrecognized regions “bronze medal”: library charges
JW: So, I figure we are on track for bronze, silver isn’t much better than bronze, so we should look at getting graph nets in to see if we can have gold SB: I agree with all the things you both said, one issue we can evade with graphnets is conformer dependence we now have. it is more in the area of “we need a solid study to weed out kinks” and not unproven science I guess DM: Lily has looked into it but there are some remaining things she couldn’t look at like how charge differences affect properties evaluated with FF
DC: I agree with all, as we approach rosemary we need to have very good torsion parameters, what I propose is a set of dihedral scans with dipeptides and feed into training. Electrostatics that aren’t aware of our specifically targeted peptide torsion scans will cause trouble MG: the charges should be baked into the torsions, sicne they’ll be known before training the torsion parameters. Also, we can make sure that the peptide conformations in the graph net training. SB – Idea would be to train net to predict AM1 charges. MG – My company has been using vchrge for a long time and it’s been fine/no major deficincies. I also wonder whether there’s a lighter-weight way to handle the pytorch/ML infrastructure since the vast majority of users won’t need to do training. CB – I like the vcharge approach. So I’d see the net as predicting electronegativities and hardnesses, which a vcharge-like approach would use. But I wonder whether there’d be a simpler way to assign electronegativities and hardnesses than using a whole opaque neural net.
CB – I like the idea of a graph net, but I’m also afraid fo some things. The thing I like is that a graph net can do anything. But the same is true of force fields - just put in the right number of terms and we can predict anything - but in practice we’ve found that making a general FF is enormously hard. It’s not clear that opaque graph nets would be able to solve this problem better than we could with automated training and manual intervention. SB – Re :Could we learn electronegativites and hardnesses without graph nets? In principle, yes. Could make an atomtype or SMARTS based scheme to assign these. But neural nets are way more flexible for more varied/experimental input formats. So, like, resonance-averaged bond orders and formal charges could be an input to the neural net. SB – Re: Opacity of graph nets: If we proceed cautiously and are rigorous with our testing, they could work well in reality. What we want from the charges from this network ISN’T perfect charges, it’s a description of the electron density. So right now we don’t trust raw electron density - see shortcomings with naive AM1 and RESP charges. So that’s why we have more advanced RESP schemes and BCCs, which are a form of manual intervention. Neural networks can look some number of bonds/hops away and so they should be grounded in reality. CB – Agree that raw physics-based models have deficiencies. But by embedding physics, they allow us to extrapolate into novel areas that the training set didn’t cover. But the nature of neural nets is that they don’t guarantee the ability to extrapolate. So generally I’m concerned about how NNs could go off the rails, whereas a physics-based model would have some extensible principles to extrapolate with. SB – I understand this concern. I wonder if there’s a way to have safeguards for “going off the rails”. Could have a really large training/test set, and check for outliers. Also, if we did distribute these models, we could do internal sanity checks (like, does a carbon ever wind up with a -2 charge?). JW – If we want to allow a lot of runway for graph net training, we could make “plan A” for rosemary be to use library charges. Then we’ll have plenty of time to look for edge cases/deficieincies
CB – Re librarycharges vs. graph net charges - Partial charges are a fixed form, so the idea outputs form both library charges and graph nets are the same. So library charges could save us from running a neural net. Then we could have a neural net just for regions that aren’t covered by library charges CB – We’ve been looking a lot at biphenyls with that conjugated central bond. You can imagine having a dataset that doesn’t have para-substituents. Then the net would look like it was trained fine, but it would be completely unable to handle an important region of chemical space and we wouldn’t know. CB -- There’s some discussion as to whether partial charges should reflect the reality of electron distributions, or whether they should be numbers that combine with the rest of the FF to yield the right answers. If it’s the latter, then we don’t need to be so strict about being tethered to physics. MG – 1) We trained charge to replicate ESPs. 2) I don’t see how this solves the problem that the trained network will do something bonkers. CB – It doesn’t solve this, but it provides us a remedy if things go wrong, and provides us a concrete target for what the neural net output should look like. MG – The hard thing is that, if we don’t base it on physical reality, then we have to do end-to-end evaluations of the whole FF at each training iteration. CB – To recap: If we have an objective reality that we’re attempting to recover with our charge model, then we will have a more straightforward time of training EITHER charge source - either graph net or library charges. SB – Agree, I’d think about doing graph-net-BCC, where the graph net output tries to predict,for example, AM1 charge, and is always adjusted by some BCCs to turn “AM1 charges” into “real electrostatic potential” SB – The other great thing about keeping BCCs around is that it splits the training to physical properties- The graph net can be fixed, and then the BCCs can be adjusted in each iteration DM – One advantage of the vcharge approach is … CB – If you’re fitting to the field that a bunch of AM1 charges produces, then you’re CB: To apply BCCs to a neural net seems like throwing away the use of graphnets, shouldn’t we be fitting to a final set of partial charges and look at ESPs after that? SB: It is a huge burden to train to ESPs but we can take baby steps and fit to AM1 first and then go forward including more complexity DC: I am up for trying out more ambitious stuff like fitting to ESPs. Right now we don’t fit charges to ESPs in our FF but we get them from AIM analysis. In the next few months we’re going to train a neural net to predict atoms-in-molecule charges not on thousands of molecules but on hundreds that would give us an idea on how many molecules we need. SB: Yeah JH mentioned that, just in case you want to train to ESPs the openff-recharge package will let you do it.
DM: So, taking a step back we don’t have enough data to actually decide on what the best path forward, and personnel to answer all the scientific questions we raised here, which might actually dictate on how we approach this, and bronze/silver seems to be viable for now SB: Step1 is to heavily test whatever charge model we have probably solvation free nergies, ESPs, hydration FEs, biopolymer torsions targets CC is working on.
In terms of science questions boils down to training different sets of graphnets and not let perfect be the enemy of the good.
DM: I think John mentioned YW is tied up and we need someone to train the graphnets SB: I’m trained one last weekend and working on resonance structures, I would love to get to the finish line before I leave. DM: If SB can push this then that’d be great. SB: Maybe we can discuss again in leadership call with John about YW’s availability. DM: So, small amount of remaining science and lots of testing.
CB: Based on the assumption that both models are convergent on ESP I think we can make a strategic decision on having a FF with library charges and then replace it with graphnet whenever it is ready.
This is what we did with RESP and AM1-BCC. DM: How close final graphnet charges are to library charges now, in biopolymer residues for example, or how feasible? MG: We could train repetitively until we get good charges CB: Then we can’t release rosemary until graphnets are closer to library charges SB: I think it is entirely feasible to converge to ESPs produced by library charges. With the flexibility we have with graphnets don’t have to train to AM1 charges, but may also be to amber99 charges or a set of charges we say it has to learn. JW: I think we can still release rosemary with library charges, right? Why do we have to wait for graphnet to match to library charges? CB: Then the decision is to go with library charges, the world I want is we train charge models to the objective reality of ESPs and converge to theoretical limit of 2/r. MG: Having that happen means that we have the same constraints library charges have in the graphnets too, for example carboxylates, etc., all can be tuned/adjusted which graphnets may not allow how would you reconcile that? CB: By having them both trained towards the ESPs of set of conformers for ASP and set of conformers for GLU, unless the carboxylates are different (I don’t think they’re), and both of them should converge to the same ESP.
SB: CB, that’s a great point that we should converge to ESPs, are there any other targets we should converge to, do we know how amber charges were derived but do we also need to throw in charges that also allow us to better reproduce torsion scans? CC: I think we train on interaction energies of charged side chain groups. DM: Historically, charges were fit first and torsions later. SB: Then it would be great CC: I can help out in the second stage once we have a MVP. DM: My issue is with personnel time, if SB can finish this until we can get YW’s time or someone else comes in. JW: I think there is a lot of overlap with CC’s work, even if we want very good graphnet charges in the future we want to have the peptide torsiondrives done carefully since these can go either way to train library charges or graphnets. DM: Right, because the torsions are derived from torsionscans on fragments we have to make sure that both library charges and graphnets agree well otherwise it affects the transferability of parameters from fragments to whole polymer CC: So, do we have to enforce the charges to be the same? DM: Need not enforce just make sure the ESPs are not entirely different CB: I know this, you can swap in AM1-BCC charges and replace RESP charges in AMBER even though it was trained with RESP since we converged the ESPs to RESP ones. The charge set can vary but the ESPs should match.
CC - Do ESPs for alanine in a folded protein converge to alanine dipeptide in gas phase used to train the charges? DM: That’d be out of scope, we’re asking if I take a graphnet representation of a small molecule, a piece of protein, etc., do their ESPs agree with fixed-charge model. We’re ignoring polarizability anyways. CB: Yeah, polarizability, conformation dependence and charge delocalization are not addressed in our FFs now.
MG: Let me show vcharge paper on correlation between charges on amino acids among charm22, amber94, VC2003. I mean you can get high correlation between library charges but may lead to funny electronegativities. CB, is there a threshold of quantitative agreement? CB: How close are these charges in terms of how they perform in the FFs can be addressed by how close the ESPs are and not on individual partial charges. MG: So, what you’re saying is we are not worried about how close the partial charges are wrt library charges but how close the final ESPs are CB: Excatly!! MG: I think I have it in vcharge paper Table5. CB: The highly polarized part of the FF is more essential for me that’s why I have a two-staged fit and tested lot of weight schemes. So, the polar part of the ESP must be similar between models. MG: Yeah, I think SB can look at the notes.
CB: A question for DC, with the atoms-in-molecules charges how well the ESPs are reproduced, since the ESPs are an artifact of distributed charges and how do they fare with that? DC: Yeah, slightly dependent on which AIM model and most of the latest ones are pretty decent in predicting the ESP. For one conformer, charges fit to ESP are better. But averaging across conformers, charges from ESP and charges from AIM give similar accuracy. We have tested DDEC and MBIS charges and they seem really good.
DM: I think we talked a lot on the science questions that need to be answered and right now bronze medal approach seems appropriate.
JW: One approach cap-and-charge can bump us from bronze to silver, and silver to gold needs to be discussed. CB: should we train to XTB instead of AM1?
|