 Hello, everybody. I'm John Kadera. I'm an associate member, that's what they call an associate professor at the Memorial Stone Kettering Cancer Center. I'm one of the PIs for the open force field initiative, which is the NIH funded part of the effort. And I'm super thrilled to be here because like I can't tell you how spectacularly happy I am that OMSF has really taken off and to see both open free energy and open force field and all of these other things like open fold, nucleating and taking shape and having an opportunity to interact and create a wonderful interoperable ecosystem. So I wanted to start and talk about really two questions for so this should be a bit more discussion provoking at the end two questions for this group. One of which is well smirnoff is the world's number one vodka. Is it really the world's number one parameter assignment scheme and I'm going to at least describe highly experimental approach that we've taken funded by the NIH grant for open force field that you can download and use in a content package very shortly, but is going to be on the experimental side for a while and so we think about whether it makes sense in the mainstream mainline for open force field. And the other one is, are we really missing out on the advantages that machine learning frameworks afford for optimizing force fields, and we're thinking about force balance the major central part of our infrastructure right now it's now over 15 years old something like that maybe more. So we have a lot of opportunities that I think we can exploit we'll have a, an illustration of what that affords in the s below my toolkit. So, as all of you folks know here, the smirnoff specification or smirks native open force field specification, which is really borrowing the only the tagging from smirks into smarts describes not just a single Adam, but a bond or an angle or a torsion or any kind of balance type by assigning a whole smirks string where you've tagged the Adam that you're interested in or Adams that you're interested in this case is describing a bond that has a certain chemical environment it's an industry standard. It's really great it's allowed huge compression of the chemical perception from Adam types there's many advantages that David mobile has talked about in the past. But it does present these challenges and to be a gave a great talk as well as Trevor about ways in which we can think about this I'm just going to you know set the stage by saying, it is really difficult to optimize these types because of the fact that they are mixed continuous right continuous discrete. So we've got these discrete types and we have to keep refining them in some way where we think about like how detailed should we get before we distinguish one type from another type and trying to do this continuous and discrete optimization at the same time especially starting from scratch is just really difficult Josh fast in my group spent some time on this with a versatile jump to be a says really made a way to find a way to make this work. It's very well, but it's, it's difficult to get a good converged equilibrium sample from this joint space. It's just a really difficult problem mixed discrete continuous optimization. So you watching Wang in the group along with Josh fast had decided that maybe, you know everything else is continuously optimizable is there some way to make the typing continuously optimizable as well. Thought about, what if we had a embeddable way of saying how different every part of the molecule was from other similar parts and other molecules. And so they came up with this graph convolutional net. I apologize that launching wasn't able to give this talk remotely today or to be here. He wouldn't let him back in the country right away so thankfully they've now allowed him to come back and assume his fellow position at NYU where he's also a Schmidt fellow, starting his independent career. But he and Josh came up with this really fantastic idea that we can take up a chemical graph that represents a small molecule or big molecule, or, you know, a protein conjugated to a small molecule for example, and then using message passing graph neural to come up with a vector that represents every atom. And what's interesting about this is that we have to be very gingerly careful with the HDMI cable. But what's interesting about this is that you automatically get chemical equivalences because of the equivariance nature of this graph. So that means that if you have methyl protons, they're automatically assigned the same vectorial representation of what their chemical is. And if you have nitrogens that are in slightly different environments because they're in slightly different regions that have different things bonded near them, they'll get different types. And so we we solve this problem of having discrete typing. Next is, you can then propagate to bonds at angles and portions realizing that a torsion fed in IJKL is the same as LKJI. They have to get the same type. So we use this symmetry based cooling to make sure we get a vector that uniquely describes each bond, each angle, each torsion, each atom. And then you can feed these into just a feed for neural network that ends up meeting the parameters that you need for your force field. And by doing that, you make a very modular approach that is also fully end to end differentiable. So if we wanted to add for point polarizability, if we wanted to add special one for parameters we just take the appropriate type coming in from the stage two. And then we add on a neural network module to predict what the parameters are for that new type and then we just refit the model. And it's fully end to end differentiable. So it learns both the typing, the type assignment in stage one and the parameter prediction and interpolation in the final stage. And you can link it up to whatever likelihood functions you might want. So right now in the open force field force balance, we use both quantum chemical targets, which are various things that look at the loss or deviation from a quantum chemical property, like energies for confirmation, as well as physical properties like, you know, equilibrium densities or dielectric constants are possibly even free energies. So we're focusing now just as this proof of concept on quantum chemical confirmations and energies but I'll show you just as an example right this can actually recapitulate what we think about as typing so if we think about just the first stage and I'll take these embeddings and then cluster them into or put them into a bin, and just to assign, can I learn a gaffe atom type right. There's a bunch of gaffe atom types. There's a lot of them. In fact, there's a lot of carbons in particular, and it does miss assign some of these, but they're miss assigning in a way that I would have trouble assigning the specific ones if you gave me the description so in this case it's sometimes confusing NF with NC, which is a inner sp2 nitrogen and conjugated system, identical to any but in a not in a non pure aromatic system. So I wouldn't know exactly which one to assign in this case there's a few other cases that that are like this as well where the confusions mimic, perhaps the, the lack of crispness in the definition of chemical environments in these. Human like confusion in those, but rest assured if you give it gaffe energies or open force field energies, it can actually learn them quite well so now I'm showing it confirmations and asking it to learn the energies and in fact it recapitulates the bonds angles torsions etc. Very well from just showing it a force field so this is also in a way something that could help us take legacy force fields like a charm small molecule force field for example for which there's not a good free parameterization engine and learn how to apply charm like parameters if we wanted to. But what's really exciting is just showing it quantum chemical directly, like we do with force balance, and then having it learn the force field directly and I'm just going to walk you through a few experiments we did to try to understand how it works and what how it fails. So, Falcato is this wonderful set that Christopher Bailey put forward, because it's very exhaustively exploring the kinds of environments you might find. If you've, if you've enumerated pretty much all of the possible environments for a simple non complex carbon hydrogen oxygen system so it has fennels alkanes ethers and alcohols. And what's nice here is that so we've broken out, we have a training set, we have a test set from this space, and we're reporting here. All of these are test molecules on these other force fields so it seems to do in terms of rms error for these different confirmations error to quantum chemistry reasonably well in terms of K calls for mole. In fact it does a little bit better than open force field does in this pretty un complex case. These are 95% confidence intervals on the upper right and lower upper lower left, lower right and upper right. Now if we show it the training set for open force field 1.2. What's surprising is that it does much better on on the test set here than it does on the on open force field does on the training set so it's a diverse set of molecules but maybe this is suggesting that somehow by allowing us continuous interpolation between different specific types, rather than being pinned to specific smirnoff types, we're actually able to increase the accuracy over what open force field did on its own training set. This is something that we've been trying for a while with these vibrant bond order based interpolations there's good physical reasons why that should happen. And maybe this is allowing us to implicitly learn some of that and how to actually deploy it. We were also suggested by Katrina Meyer, the vehicle data set which is this really exhaustive exploration of bicycle bicyclic scaffold heterocyclic scaffolds of the future, which could be important for future compounds that might be. Give us slightly different bond vectors and slightly different polarities inside of binding sites, some of these look a little crazy obviously with the number of nitrogens you find in them, I wouldn't want to be caught dead anywhere near them because I might end up dead. But what was surprising is how much better as below my did in this case, then the open force field and in fact all the other models. The reason actually turns out to be quite simple so these were all subjected to QC fractal quantum chemical minimizations. And it turns out that some of these were not aromatic at all there was a cheminformatics error in their preparation, where they simply didn't satisfy aromaticity, and the quantum chemistry picked up on this and correctly made these weird nitrogen containing compounds appropriately pyramidal. And I don't know how realistic it would be for a synthetic chemist to access these scaffolds, but the fact that just by having a couple of examples sneak into the training set. Espeloma was able to pick on pick up on the fact that these should be different types, and then assign them different types was actually really heartening that suggests to us that it's really able. It gives us a new paradigm for solving the problem by just coming up with more examples, rather than having to have someone like Christopher help nominate new types. We can also show peptides, and it seems to do better than amber 14 sp does, which is terrifying, because a lot of time and effort has gone into parameterizing amber over the years and fixing a number of issues. We can in fact show it both peptides and small molecules and come up with a joint small molecule peptide force field so this is the first combined force fields that I think the effort produced until Chapin's new version comes out of a self consistent treatment, but it does seem to show stable simulations which is really impressive for something that is still using Leonard Jones and a and one BCC from open force field 1.2 in this case, but then just re typing all the valence terms now of course there's not a big opportunity for it to go off the rails right it's just learning the valence types in this case. What's really cool about this is that it does display the self self consistent size properties that you might expect. There's a certain limit to the number of rounds of message passing that you take. So you can't have effects from the ends extend far too far. So if you're looking at inserting different numbers of alanine inside of a cap alanine peptide, all of the online and all the charges, for example, as we plot here, all seem to stabilize after edge effects in position which may be realistic right and then seem to be pretty stable so that you can, you know, you can actually type full biocall race and actually very fast, there's some catastrophe here on the right side but you'll see that everything takes less than a second even up to 500 residues which is pretty impressive. And that's just on a CPU. In fact, you can take much less than that on a GPU. So you can actually use a GPU to assign parameters, perhaps for many things at once because you can bundle all of the graphs into a single large graph that doesn't have connections in it so if you need to parameterize many molecules. The really cool thing I thought was that who valent ligands are now breeze right by have a brute nib in its pre reacted form this is a covalent kindness inhibitor of BTK. It has this warhead over here that changes when it actually binds to the system here's the peptide part of BTK that gets changed by or labeled by this root nib. So you can see that if you it's hard to see with the small numbers here but you can take my word for it that there's some rearrangement and charges and bonds right around the reactive group, but the rest of the molecule remains unperture. So it doesn't do catastrophic things to your molecule. And in fact this will, we can also self consistently assign charges using the same kind of approach that Lily spoke about I think today, this morning, yesterday. We'll speak about did or will yes in the keynote. So what's cool here is that okay we use that this to apply small molecule parameters because we are now only only now thanks to the bio polymer support for PDV reading from open force field able to get the proteins in for into our tool chain for energy calculations. If you just replace the small molecule parameters, it seems to do reasonably well. It's very competitive with open force field 2.0, maybe slightly better, but it's hard to say. You need to try more systems to be able to say with certainty or confidence that it does, does any better than open force field 2.0 2.1. But the other cool thing is you can also propagate just like we see with other. We saw with host guest systems you can also propagate free energy so anything you can compute and differentiate you can also include as a target. And in fact the differentiation is automatic since we're using machine learning framework I'll get to that more in just a moment. I've mentioned the, there's, there's, there, there are potential advantages to using both the conformer energies and partial charges or potential ESPs in training a joint model because it seems that there's information content about learning about differences in in atom chemical environments by training everything together so you do see a reduction in error. If you train on both the energetics and the charges at the same time, because that's another important thing to keep in mind. I've already shown you stable simulations. So the next thing was, this was a very small training set. I believe we were limited in the size of the training set data initially for parsley because of the lack of scalability of force balance getting it to feed it many more quantum chemical data measurements. Also, we just didn't have a lot of data in QC fractal at the time that was appropriate. So as part of a collaboration between open force field and open, and open MM, I apologies I don't have all of the names here but you should go read the paper, or check out the data set for all of the folks that contributed it was a really a team effort generated a very large multimillion snapshot data set that has very good elemental coverage that provides us with information about the chemistry the chemical compounds that we have including their bonds that went into this so we can parameterize from that information. And he doesn't have any bond information which makes it a very difficult set to use plus it's also very limited in in elements. This set was generated with both a very high level of theory and a lower level of theory or the lower level of theory that open force field uses as its default specifications that be mixed and matched. This was generated slightly differently from what we've done with open force field in the past, because unlike the optimization data sets which are trajectories of optimization, or the torsion data sets and I should say that all of the data I've just shown you was just using optimization data sets we didn't even use any torsions, we seem to be able to recapitulate torsions reasonably well. If we do use that torsion data that's, that's great to we use every point along the minimization trajectory, however, for the optimization data set, but these are generated with MD simulations using, I think, either gaffer open force field I can recall which, and then using off equilibrium snapshots at slightly elevated temperature to get a better representation of what the energy surface might look like. Now Ken Takaba in the group has been working on extending this to nucleic acids as well. He's particularly interested in RNA you might note of notice that there's a huge number of companies that are interested in drugging RNA with small molecules or otherwise using them as a platform. I had some RNA injected into my arm maybe you did too. So suddenly become very important. Ken has worked with folks in the open force field side like Pavan and Trevor and David Donson to help generate a large RNA data set that supplements the spice data we can do the same with DNA at some point as well. And he's been collaborating with you on training on this updated data set that includes a subset of spice that is all of the open force field generation to optimization data. And then there's the base hub chem molecules which include a lot of different kinds of chemistries, both the dipeptides and monomers from the Desrez set, and the also the dipeptide, a dipeptide representative MD set, as well as the RNA data. And what's really interesting is that it seems it's very hard to see these numbers but it seems to be all existing force fields at least in this energy metric. So we hold out some of the data from each of these sets to leave us test set data hold out some of the molecules, and then we use that to estimate the test data set, which is also estimated for all of these other force fields so even for RNA force fields RNAOL3 for member, it seems to do very well for energetic properties. They're calling it spicy espeloma for this, but it'll be espeloma 0.3.0 or 0.4.0 depending upon where they settle is already a pre release available. I'll talk about that in a moment. Thanks to David Mobley helping us work with Pavan. We've gone through several iterations of benchmarking and identifying limitations and failures and improving things like how we treat the improper torsions. It turns out we were trying to predict too many improper torsions and there are still many things that can be refined with this perception approach about how we assign torsions because we we let it assign any sign to the case. So you can get the phase as a continuous parameter from one n equals one all the way through n equals six periodicity to torsion so we've had to trim that a little to do what open force field does to limit it to one or n equals one or n equals two firm properties, but it does seem that the quality these quality metrics are suggesting that it's doing very well on this industry nominated benchmark set. It also seems to be doing reasonably well on free energy calculations again, rivaling the open force field 2.0 in accuracy will have to see how that works, how that plays out for many more, many more sets. Okay, so that's the first question or proposal is that I think this might be a competitive way of solving these challenges in mixed continuous discrete optimization. And that's something that that we could we can investigate as a potentially forward in parallel to Smirnoff for a while but possibly to replace it sometime in the future. The other thing is, I wanted to engage with you about how useful it has been to use these machine learning frameworks there's a lot of effort behind these a lot of folks are using them. This is now a very old example but it shows you just how exciting it is to have, you know, somebody has curated a standardized data set like the open force field has and you can access that programmatically in one line. You have a novel architecture that you can define which might be a completely new type of architecture using powerful abstractions that thanks to the toolkit writers, you can compose very easily. And then using best practices standard best practices you can optimize it assess how well it's doing and then actually use it just in a few lines of code. And we can do similar things with force fields if we really really wanted to. This is not a working example but certainly the inspiration for what we'd like to do what if I just comment out the bond charge corrections or add in a point polarizability or add in a special one for exception and then refit and then do these experiments very easily. This is a working example of that realization this is fitting an entire force field in espeloma. It's very easy. It's also very easy to extend because I just need to add a few more lines to express whatever energy functions I want to implement. It makes force field science very easy in terms of exploring different models, at least if you're doing fitting to quantum chemistry takes more effort of course to fit to experimental data. Now, these machine learning frameworks are great because first of all, another major major companies are pouring money into maintaining them so we don't have to that relieves us of some infrastructure burden right we have to maintain, along with the The other thing is hardware acceleration is baked into these machine learning learning frameworks and they're always investing in making them faster. We could build force fields on GPUs as a result. They're scalable these force field these frameworks have scaled some of the largest hardware installations available, probably much larger than we have access to. There's a lot of libraries of optimizers or Bayesian samplers or tools for tracking the experiments that you run, like weights and biases. There's also a lot of support, there's many people using this compared to the number of people using force balance. So, there are two experimental versions of espeloma available right now. They're almost condit installable. Right now we're working on one of the tool chain issues I'll get to in just a moment. One is a pytorch based version, and the other is a jackspace version, the pytorch one is much more mature the jacks is even more doubly experimental, but they fundamentally do the same thing. It's an exciting approach because I think many of you know pytorch but jacks is combining Python with autograd, and then the Google based back end that TensorFlow uses called XLA, which translates code that's been compiled just in time to run very fast on hardware like a GPU or a TPU or a CPU or any new hardware that comes along that you might want to actually take advantage of somebody else is going to take advantage. I was going to figure out how to do that quickly, and it just is Python with four superpowers one of them is jit which is make it fast, you can wrap any function in that rad which is give me the gradient with respect to parameters or whatever. P map which is do it in parallel on the GPU or TPU, and then P map is do it in parallel across many nodes and you can combine them in any way you want. And it's been the ML choice of deep mind for a long time and likely will replace TensorFlow at Google if things keep going the way they're going. So if we think about what impact this would have for our communities right the scientists might be able to rapidly carry out faster refitting experiments if they could fit with the GPU. And it's much easier to explore new potential terms because it's much easier to code one of those up than the whole plug in that takes advantage of something in OpenMM. They can experiment with these potential with ML potentials as well because they're all written for these frameworks and not for OpenMM can simultaneously build and assess both Smirnoff and Espeloma style force fields would be the dream so you can optimize parameters fixed Espeloma types or you can optimize an entire Espeloma set or you can even build these Bayesian ensembles and explore Bayesian force field parameters of families for the engineers who have to do this for big fits, it would be fast and scalable and reproducible. And for industry, I'd like to think about this paradigm, you know, could we make it easy enough for people to use an industry that we are producing the foundation models that people take home and tailor to the data sets and problems they have internally. This could be a new paradigm for industry to really rapidly fine tune and get better results out of our force fields for the problems that they care about. And we haven't focused on making it easy to use for industry yet it's possible, but if we were able to do that it would be much easier with these deployable machine learning frameworks. And there's a few opportunities as well and about how we could how we could overcome some issues to make this better one of which is that we can use our existing tools like the property calculator that Simon and others have engineered for computing free energies and gradients. A lot of it is setting up workflows to build these systems and simulate them in a way that is automatable. But at some point we can integrate these ML implementations of the energy functions to evaluate the gradients much more efficiently. We can also potentially ultimately replace the open MM layer with a ML based simulation engine if one comes along. There's a few experimental ones like torch MD, and a few others at the moment but this might be something that is also an opportunity in the future to let us take better advantage of these big GPUs which waste a lot of time by simulating a very tiny system pack them all together. David's Rudy of course has another way to do that but this would be one one possible alternative. And of course condo packaging is a challenge because this community has standardized on tip. Now, that's a discussion that we can have that's much larger but there's a lot of opportunity for us to help work with these communities and get all of their tools on to condo as well. We've managed to do this for everything except DGL which is the final piece because it has lots of dependencies that have been pulled in. So we're almost there and Mike, Mike Henry has been really pushing this forward so we're almost to the point where you can install it like Honda, you can very easily install it other ways though. And so, for if there's any time left, I'd also be interested in what folks think about, you know, we do have some opportunities and we have to figure some things out one of them is this benchmarking suite that was announced yesterday is going to be really exciting with the modules that allows to plug in different ways to compute data and compare against it the same modules are used here to parameterize and we would want to compute ratings of them so there's an opportunity to avoid duplicating our work by architecting these properly so that they can go into either framework perhaps. We also want to think about what is the MVP what minimum functionality would we need in order to be able to swap out force balance at least for some tasks would it be interesting. We could try the industry tailoring, for example, where you they can easily generate quantum chemical data or any like data that they can use to tailor their quantum chemical to generate quantum chemical data sets to tailor just the balance, but stick to our non bonded for example that could be an easy win, would it be really useful to do balance fitting to quantum chemistry for the scientists for force open force field to be able to try out new functional forms very easily. And we have to think about what that MVP is and then is there a path to deprecating force balance is it something that we want to keep alive all the time we don't want to have to support two different routes to do something forever, but there might be some sort of sensible rap to thinking about sunsetting it in the future. And with that, I'd like to thank all of you for your attention and if you had any thoughts on the discussion questions I'd welcome them during the questions.