Open Source Alchemy with OpenFE: Pacifichem 2025

1 Open Free Energy Open Source Alchemy with OpenFE Pacifichem 2025 Irfan Alibay, OpenFE Science & Technical lead Download this presentation 2 Open Free Energy Calculating free energies is key to understanding biological processes Image credit: John Chodera Driving affinity and potency Partition coeffs and permeability Predicting drug resistance Structure-enabled ADME/Tox targets Predicting crystal polymorphs, solubility Driving selectivity 3 Open Free Energy Rigorous free energies via alchemistry Molecule BMolecule A λ=0.5 λ=1λ=0 Hλ) = 1-λ HA + λ HB Transformation Path Slow FastFast Slow 4 Open Free Energy Why do we need OpenFE after decades of successful alchemy? ●Increase access and ease of use. ●Encourage best practices. ●Reduce duplication of efforts. ●Create a collaborative environment which encourages and supports future scientific developments. A need for open and free alchemical tooling Free Permissive open source MIT) license Easy to Use Designed for simple, large scale deployment Accurate Guide pharmaceutical drug design and discovery Extensible Clear API to allow reuse and adaptation of components Robust Built and maintained to industry standards 5 Open Free Energy The OpenFE core development team Josh Horton Alyssa Travitz James Eastwood Hannah Baumann Irfan Alibay David Mobley Iván Pulido Ian Kenney Project Staff David Dotson Mike Henry Project Director Chodera Lab 6 Open Free Energy OpenFE Supporting Partners Technical Advisory Committee Oliver Beckstein Phil Biggin John Chodera Zoe Cournia Sukrit Singh Industry Partner Members 18 companies David Mobley Bharath Ramsundar Michael Shirts Jonah Vilseck Emilio Gallicchio Achira Amgen AstraZeneca Bayer Biogen Boehringer Ingelheim Cresset Deep Origin Eli Lilly Genentech (Roche Group) GSK Antonia Mey Julien Michel Stefan Boresch Sereina Riniker Bristol Myers Squibb Charm Therapeutics Confo Therapeutics Johnson & Johnson Merck KGaA Neomorph Pfizer 14 academic members 7 Open Free Energy Part of the Open Molecular Software Foundation Open Force Field Open Free Energy openforcefield.org openfree.energy OpenFold openfold.io WESTPA westpa.github.io OpenADMET openadmet.org 8 Open Free Energy The OpenFE ecosystem gufe custom data models konnektor network construction LOMAP atom mapping (originating from the Mobley Lab) kartograf atom mapping focused on 3D geometries cinnabar Free energy networks analysis and best practices alchemiscale distributed execution across heterogeneous compute openfe central framework and entrypoint for FE calculations feflow Community Protocols and many more! exorcist Local HPC execution workflow openfe protocols Alchemical Protocols pontibus Experimental Protocols for FF development 9 Open Free Energy Supported by a common Python API for alchemistry gufe custom data models AtomMapper LOMAP Kartograf GUFE Abstract layer Application Layer GUFE is our common language for free energy simulations. Defines various code building blocks for creating new tools and methods. 16 Open Free Energy Pontibus: experimental alchemical Protocols for FF development INFRASTRUCTURE Initial Force Field FF & experimental conditions Computed properties Parameter assignment Force fields, molecule sets Experimental Property Data DATA Optimized Force Field Benchmarks ForceBalance regularized least squares Open Force Field Release SFEs of MNSOL evaluation set SFEs of freesolv evaluation set Lily Wang 17 Open Free Energy Pontibus: experimental alchemical Protocols for FF development Benchmarking the new OpenFF 3.0.0a protein force field alpha candidate. Jeff Wagner Matt Thompson Chapin Cavender 18 Open Free Energy FEFlow: Community-developed Protocols Iván Pulido 19 Open Free Energy FEFlow: Community-developed Protocols 19 Iván Pulido T2A is a known difficult mutation Implementing protein mutations for hybrid topology relative free energies based in FEFow. Zhang I., Singh, S., et al. 2023. J. Chem. Theory Comput. 19, 15. 20 Open Free Energy Konnektor a library for building alchemical transformation networks. Kartograf a 3D-first atom mapping utility. Simulation setup tooling: Konnektor and Kartograf Ries et al., JCTC, 2024 Star Network Minimal Spanning Tree Network Starry Sky Network Ries et al., JCIM, 2025 21 Open Free Energy Cinnabar (formerly arsenic) network-wide alchemical analysis. openfe-analysis tools for post-simulation analysis. Simulation analysis tooling Post-simulation analyses with openfe-analysis Cinnabar (formerly arsenic) 22 Open Free Energy Large-scale execution via Alchemiscale Ian Kenney David Dotson Alchemical Network AlchemiscaleClient AlchemiscaleAPI State Store (Neo4j) Object Store (AWS S3) AlchemiscaleComputeAPI ComputeService FAHComputeService ComputeService alchemiscale serveruser workstation Kubernetes cluster HPC cluster Folding@Home work server Strategy Result Result Result User 23 Open Free Energy Large-scale execution via Alchemiscale 23 Ian Kenney David Dotson Over 200,000 free energy calculations, totaling 1.2 million GPU hours over the last two years. 24 Open Free Energy Folding@Home support with Alchemiscale Task Index SetupUnit Transformation ResultUnit Process Pool SimUnitSimUnitSimUnit FAHComputeService Folding@Home work server Ian Kenney David Dotson A single FAHComputeService can execute thousands of ProtocolDAGs concurrently, ensuring Folding@Home is saturated whenever Tasks are available to compute. Hugo 🦘 ! (ASAP Discovery) F@H yields the same results as traditional HPC. 25 Open Free Energy So are we ready for production use? Assess performance Get production ready Start using OpenFE in production Get people familiar with OpenFE A large-scale industry benchmark initiative to evaluate OpenFE. 32 Open Free Energy Private dataset results Pairwise absolute ΔΔG error Deviation and Correlation Metrics 33 Open Free Energy Private dataset results 34 Open Free Energy OpenFEʼs default performance is competitive with state-of-the-art tooling FEP. Application to real world systems (i.e. private datasets) demonstrates reduced accuracy, but retains predictive power on most submitted systems. Our industry benchmarking results offer a wealth of feedback to help us define future developments. Benchmark conclusions Read our preprint on chemrxiv! 35 Open Free Energy What are we currently up to? 36 Open Free Energy Membrane support Image credit: Yuxuan Zhuang Improving OpenFEʼs domain of accuracy through support for membrane-bound protein systems. A2A OX2 P2Y1 A2A OX2 P2Y1 Lipid exposed binding site Ross et. al (2023)OpenFE 38 Open Free Energy Improved hybrid topology relative free energy accuracy Addressing issues identified by Fleck et al. (2021) on hybrid topology dummy atoms. Introduction of redundant terms from disappearing atoms is causing ligand behaviour at end states to differ from non-alchemical systems. Ultimately has an impact on the accuracy of relative free energies. 39 Open Free Energy Improved hybrid topology relative free energy accuracy Currently investigating different potential fixes, including term scaling and removal. 40 Open Free Energy Improving openfeʼs simulation handling infrastructure, including: ●Protocol continuations & extensions ●Better HPC submission & job monitoring ●Efficient and standardized output file structures Improved execution & usability OMSF eco-infra David Swenson Ethan Holz 41 Open Free Energy Where does machine learning fit in? 48 Open Free Energy AshGC: Graph Neural Net for Partial Charges Sage 2.3.0 performs comparably over JACS RBFEs. 49 Open Free Energy Binding free energies from OpenFold3 co-folded models Can we use OpenFold3 predicted structures as inputs to OpenFE? Read about the OpenFold3 Preview here! 50 Open Free Energy Binding free energies from OpenFold3 co-folded models Initial feasibility assessment using ASAP-discovery SARSCov2 Mpro series. Caveats: ●Mpro structure likely included in training set. ●Ligands from other scaffolds could have been included Jenke Scheen 51 Open Free Energy Binding free energies from OpenFold3 co-folded models Docked OpenFold3 Read our OpenFold3 to OpenFE tutorial! 52 Open Free Energy Binding free energies from OpenFold3 co-folded models Very similar initial binding poses. Grey: Crystal, Purple: OpenFold3 Bad mapping from docked poses. 53 Open Free Energy OpenFold3 to OpenFE: issues still need to be addressed Uncapped chain termini. Distorted ligand conformations Distorted ligand conformations / binding site clashes ● Run pose busters to filter through models with some manual inspection Complex protonation is difficult with open source tools ● Either use paid-for tools or having to rely on re-preparing and protonating protein & ligand separately. 54 Open Free Energy DragonFold-FEP Work by Charmm Tx [Scheen et al. (ChemRxiv)] demonstrates that co-folding models can be useful alongside FEP / OpenFE. 55 Open Free Energy Fine-tuning affinity prediction models Investigating how OpenFE results can be used to fine-tune ML models. Initial tests fine-tuning FLOWR.root Cremer et al.) to improve affinity predictions for MPro set. Base modelFine tuned All data Train set Test set 56 Open Free Energy Future directions in the ML/FEP space we are excited about! ●Continued improvements to integration with co-folding models ○Fully automated OpenFold to OpenFE? ○Data augmentation for affinity training? ●Direct use of MLIPs in alchemical simulations ○Alchemical transfer method (e.g. QuantumBind-RBFE ○End state corrections (e.g. Rufa et al) ○Direct ML/MM FEP using 4D decoupling (e.g Picha et al.) ●Active learning ○Large-scale campaigns using Folding@Home! ●Force field parameter generation & refinement 57 Open Free Energy Thank you! OpenFE developers Hannah Baumann Josh Horton Alyssa Travitz Mike Henry James Eastwood Ivan Pulido David Dotson Ian Kenney Benjamin Ries David Mobley Contributors in the benchmarking study Abbvie ASAP Discovery AstraZeneca Bayer Biogen BMS Boehringer Ingelheim Deep Origin Eli Lilly Genentech GSK Johnson & Johnson Merck KGaA Odyssey Roche Chodera lab & ASAP Discovery John Chodera Dominic Rufa Ivy Zhang Jenke Scheen Hugo MacDermott-Opeskin OpenFE TAC & Board OpenFold Jennifer Wei OpenFF Lily Wang Jeff Wagner Matt Thompson Chapin Cavender OpenFold