April 2026 Infrastructure Update

Hi everyone. I'm Jeff Wagner, technical lead at OpenForceField, and this is our first quarter 2026 infrastructure update. We've had a number of minor releases to our major packages. We've had two releases of Interchange that focused on bug fixes and minor improvements. We had a minor release of Nagle that improved support for custom toolkit registry usage, which some power users might want. We made a new release of Pablo that fixes a couple bugs that we found in internal testing on some polymers and modified proteins, as well as adding a new function to create new residues with additional crosslinks. We've also been coordinating with OpenFE to help coordinate maintenance and updates to Pontibus, our tool that kind of sits between the projects. Uh this is largely driven by Erfan Alabai and Alissa Trappett at OpenFE, but we're we're figuring out how OpenFF can help support the maintenance and development of this project more. Something that you should see soon is a significant update to Smirnoff template generator and the OpenMM Force Field ecosystem. They will be gaining the ability to parameterize proteins using Smirnoff force fields as well as assign virtual sites from Smirnoff force fields. This effort was largely driven by Evan Pretty, an OpenMM developer, with support from Peter Eastman. Uh in the next OpenMM Force Field and OpenMM releases, uh we'll have support for multi-residue molecules, so basically proteins, parameterized using Smirnoff force fields in OpenMM Force Field, as well as virtual sites. Also, over discussion of how these things have been used in practice, we defined some new behaviors. Some of these are breaking, but they sort of have the full weight of OpenFF and OpenFE and OpenMM behind them, so feel free to talk to us if you don't like these, but we're pretty sure these are improvements. Uh one of them is Smirnoff template generator now will always require definition of which force field you intend to use. Previously, this wasn't the case and people would do studies uh using a default argument and they had no idea which Smirnoff force field was actually used. OpenMM force fields now will honor constraints from Smirnoff sources. Previously, it removed all constraints and then added them based on a keyword in the OpenMM API about which constraints the users wanted to add. Now, constraints from Smirnoff sources will always be in the final OpenMM system and the OpenMM API keyword can be used to add additional constraints. Evan also found that this was unsuitable for proteins because of some scaling issues for larger uh molecules and so Evan did some major improvements that provide uh faster improvement uh faster performance on large molecules like proteins and biopolymers. We're happy to announce that we concluded the 2026 virtual workshop series. We had two unique workshops that we ran. One was on simulating proteins with PTMs as well as highlighting a number of other best simulation best practices. And one on fitting a Smirnoff force field using PyTorch, which reflects the new fitting stack that we're going to be moving towards as we uh try to put force balance behind us. The workshop materials for both of these are available at the GitHub repository linked here and I'll put this in the chat during the meeting. The fitting recording will be uploaded soon. The recording of the PTM workshop is already uploaded. This is thanks to the SBGrid webinar series hosted by Harvard. Uh you can watch us on YouTube. This actually came out great and I really strongly recommend that people watch it. This is also the first time that we've made environments available using both conda and pixi. Uh pixi is a tool that can install that can basically create environments including conda packages and pipi packages. Uh it's very fast and it a lot of people are moving over to it. So over here I'm showing what a pixi.toml file looks like. This is how it defines dependencies for a core of a package and then dependencies for optional features of a package. I just wanted to to show you all this because this looks a lot like a conda file. It's not exactly the same as conda yaml but it is a pretty close match with a a little work. Um so if you hear your colleagues in the future talking about using pixi, this is just what it looks like. It's a lot like conda. We will continue providing all of our instructions and stuff um assuming that people use conda for the the foreseeable future. But just to let you know, for some cases we will additionally be providing pixi instructions. I also wanted to point out that in the PTM workshop, we show an example of how to combine a rosemary protein with amber lipids. This is something that we we hadn't shown very much before but I thought that I should highlight it. Here we're constructing an OpenFF force field using the rosemary alpha and the OPC3 water model and we're creating an interchange for that. But additionally, we're using the OpenMM force field class to load uh a port of an amber lipid force field and we're using that to parameterize the lipid component here and then finally we're making the simulation system by using the interchange.combine method. This is something that's that's good enough for us to put in a workshop. If you want to use membranes in your workflow and you've been waiting for the go-ahead from OpenFF, uh consider this your go-ahead. You can use basically exactly this code and you should be able to to set up and run membrane simulations. Note that we don't build membranes, so you're still you still need to come to us with coordinates and your and your box. But we can parameterize them once you have it built. Finally, uh the the paths of fate and employee turnover or turnover at OpenFF have left us with uh 2.8 industry-funded FTEs or like core-funded FTEs on the infrastructure team and just one FTE on the science team, that's Lily. We also have Jennifer Clark, but she's funded on a special project. This is not great because our mission is to deliver you better force fields and not just maintain software. So, we are rebalancing Matt Thompson and Ashley Mitchell's efforts towards infrastructure supportive science goals. Matt's working on supporting uh our force balance replacement. Our PyTorch-based based fitting stack is already quite good at some valence fitting tasks, but it can't yet do fitting to uh physical properties. Matt's implementing that. Ashley's working on infrastructure to benchmark Rosemary candidates on cyclic peptides. This is being done in conjunction with Chapin. Uh we have a Kubernetes-based cluster that can provide us more GPU compute. That would accelerate our protein force field efforts, but we need to kind of rewrite things to run on on Kubernetes efficiently with the GPUs that they have available. So, that's our update for Q1. Hopefully, I will have uh some science progress to show you with infrastructure effort at the end of Q2. Thank you.