Overview

Understanding the essence of life requires an integrated approach that spans the vast spatiotemporal scales from atoms and molecules to entire cells. This workshop aims to establish the foundational methodology for “AI-driven multiscale simulation” that harnesses recent advances in AI technology to bridge simulations across different hierarchical levels.

Through this workshop, we seek to create a new research paradigm for achieving a truly integrated understanding of life phenomena from molecules to cells, establishing “AI-driven Multiscale Life Science”.

Key Topics

Registration

Registration is now open. There is no registration fee.

Both on-site and Zoom participation are available. On-site registration will close on April 30, 2026 (Thu) or once capacity of the meeting room is reached (maximum 100 people), whichever comes first. Poster presentation slots (maximum 40) are also available upon registration. We welcome poster presentations from all participants, and particularly encourage early-career researchers and graduate students to present.

Register Here

Venue

8th Floor, Integrated Innovation Building (IIB), RIKEN

6-7-1 Minatojima-minamimachi, Chuo-ku, Kobe 650-0047, Japan

The nearest station is Iryo Center (Medical Center) on the Port Liner. From the station, follow the 2F pedestrian deck directly to the IIB building and enter through the 2F entrance (there is no 1F entrance).

This workshop is a hybrid event. All talks will be streamed via Zoom for remote participants. The Zoom link will be shared with registered participants before the workshop.

Lunch information: There are limited dining options near the venue. Lunch boxes will be provided for invited speakers. Other participants are encouraged to bring lunch or use nearby convenience stores. Please refer to the nearby restaurant & convenience store map (PDF) for available options around the venue.

Program (Tentative)

Each talk: 25 min presentation + 5 min Q&A = 30 min

Click on a talk title (▶) to view the abstract

Day 1: May 28, 2026 (Thu)

08:55 – 09:00 Opening Remarks — Yuji Sugita (RIKEN / University of Tokyo)
Session 1: AI/ML for Molecular Dynamics — Sampling, Kinetics & Transitions (Chair: Yasuhiro Matsunaga)
09:00 – 09:30 Pratyush Tiwary (University of Maryland)
Efficient, explainable and extrapolative AI for biophysics
09:30 – 10:00 Xuhui Huang (University of Wisconsin-Madison)
Machine Learning Models for Non-Markovian Protein Dynamics
10:00 – 10:30 Chris Chipot (CNRS)
From Atoms to Pathways: A Machine Learning Perspective on Molecular Transitions
Understanding rare events in complex molecular systems requires identifying transition pathways that are both physically meaningful and committor-consistent. I will introduce a neural-network-based framework that simultaneously learns the committor function and the associated committor-consistent transition pathways, providing a unified, data-driven perspective on molecular transformations. The method, rooted in the committor time-correlation function, operates beyond the overdamped diffusive limit and distinguishes multiple competing mechanisms. Next, leveraging geometric graph neural networks built on vector perceptrons, I will show how the committor can be learned directly from atomic coordinates—bypassing the need for predefined collective variables—and yields atom-level interpretability by pinpointing the key atomic contributors to the transition. Last, I will present an iterative refinement between biased sampling and committor learning to identify dominant transition channels, enhance sampling efficiency, and enable accurate estimation of rate constants. Demonstrations on benchmark potentials and molecular systems, from peptide isomerization and folding models to chemical reactions, showcase the robustness, transferability, and ability to uncover physically grounded, of these committor-consistent descriptions of complex molecular processes.
10:30 – 11:00 Anand Srivastava (Indian Institute of Science)
Morpheus: Investigating fold switching proteins through the lens of machine learning and statistical thermodynamics
Functionally important metamorphic proteins, which can switch folds between different well-folded conformations, are now thought to be widespread. These protein sequences do not obey the classical folding dogma, and the underlying principles that drive their fold-switching behavior is poorly understood. In my talk, I will present two diverse approaches towards understanding the molecular grammar of fold switchers - one based on machine learning and other on thermodynamics principles. In the first part, I will present a new fragment-based classification algorithm from my group, named as Morpheus, which can annotate fold-switching proteins from sequence information across proteomes (https://doi.org/10.1093/bioinformatics/btaf635). Since Morpheus server takes in user defined sequences, it can also act a useful screening tool towards the de novo design and engineering of such proteins through further experimentation. However, rationale de novo design of fold switching proteins will require a thorough understanding of the molecular driving forces leading to their conformational propensities. In the second part of my talk, I will discuss our recent unpublished work (https://doi.org/10.1101/2025.10.10.679401) on understanding the thermodynamic design principles that leads to fold-switching behavior in such proteins. As a test case, we considered a recently designed fold-switching protein [John Orban and co-authors, PNAS 2023] that transitions between a 3α fold and α/β fold upon changes in temperature. We performed a detailed thermodynamic analysis on this protein using an advanced and efficient free energy all-atom molecular simulation approach. We find that while 3α fold is stabilized at low temperatures by enthalpic contributions from favorable water-water and protein-water interactions, the transition to the α/β fold at high temperatures is driven by the gain of entropy from the release of ordered water molecules surrounding the 3α conformer into the bulk. Our study elucidates the importance of thinking in terms of entropy-based design in natural systems and provides a integrated machine learning and physics-based framework that can help in the design and engineering of future synthetic and functional metamorphic proteins.
11:00 – 11:20 Coffee Break
Session 2: Protein Design, Engineering & Drug Discovery (Chair: TBA)
11:20 – 11:50 Juyong Lee (Seoul National University)
TBA
11:50 – 12:20 Takahiro Kosugi (Institute for Molecular Science)
Computational Protein Design for Rational Control of ATPases
ATP hydrolysis plays key roles in the functions of numerous proteins, including molecular motors and kinases. Therefore, the ability to control ATP hydrolysis and the resulting ATPase-dependent functions would not only deepen our understanding of biological systems but also enable their targeted manipulation. Recent advances in computational protein design have been substantial, and we recently achieved the de novo design of an ATPase (T. Kosugi, et al. Protein Sci. 2025). Building on this technology, we also introduced artificial allosteric sites into a rotary molecular motor, resulting in enhanced rotational speed (T. Kosugi, et al. Nat. Chem. 2023).
12:20 – 12:50 Jonathan Essex (University of Southampton)
Engineering antibody structure and function by multiscale simulations
12:50 – 14:10 Lunch
Session 3: Membrane Proteins, Ion Channels & Molecular Machines (Chair: TBA)
14:10 – 14:40 Syma Khalid (University of Oxford)
Multiscale simulations of the E. coli outer membrane
The outer membrane of E. coli a proteolipid bilayer - new details of its spatial organisation are emerging from molecular imaging, chemical biology and molecular simulations.

We use multiscale molecular dynamics simulations to explore these details at a resolution that is not available to experimental methods. A largest, coarse-grained simulations are composed of over 10 million particles. Fine grained resolution is added with smaller simulations at all atom resolution. We also explore the impact of antimicrobial agents on these membranes.

14:40 – 15:10 Yi Wang (Chinese University of Hong Kong)
Functional Structure of the Transmembrane Channel from the Vacuolating Toxin of H. pylori
15:10 – 15:40 Kei-ichi Okazaki (Institute for Molecular Science)
TBA
15:40 – 16:10 Jejoong Yoo (KIAS)
Overcoming Biomolecular Modeling Challenges: AI Reconstruction and Accurate Simulation of IDPs and DNA-Binding Proteins
Simulating intrinsically disordered proteins (IDPs) and DNA-binding proteins is challenging because their highly flexible regions are systematically missing from experimental structures. To create complete, simulation-ready models, we introduce PATCHR, an AI-powered diffusion model that accurately inpaints these missing segments. However, even with complete structures, standard molecular dynamics (MD) force fields often overestimate non-bonded attractions, which artificially stalls physical movements during simulations. To resolve this, we present CUFIX-AMBER, an improved force field with carefully calibrated nonbonded and hydrogen bond parameters. Together, this toolkit bridges generative AI and physical simulation to successfully capture the true conformational ensembles of IDPs and the physical diffusion of proteins along DNA.
16:10 – 16:40 Kei Moritsugu (Osaka Metropolitan University)
TBA
16:40 – 17:00 Coffee Break
17:00 – 18:00 Free Discussion / Poster Preparation

Day 2: May 29, 2026 (Fri)

Session 4: Multiscale Modeling & Coarse-Graining (Chair: Florence Tama)
09:00 – 09:30 Gregory Voth (University of Chicago)
Ongoing Advances in the Theory and Application of Coarse-graining
Advances in theoretical and computational methodology will be presented that are designed to simulate complex (biomolecular and other soft matter) systems across multiple length and time scales. This bottom-up approach provides a systematic connection between all-atom (AA) molecular dynamics, coarse-grained (CG) modeling, and mesoscopic phenomena. At the heart of these concepts are methods for deriving CG models from molecular structures and their underlying atomic-scale interactions. An important component of our work in the past few years has been the concept of the "ultra-coarse-grained" (UCG) model and its associated computational implementation. In the UCG approach, the CG sites or "beads" can have internal states, much like quantum mechanical states, so the UCG model involves a conceptual abstraction beyond simply Newtonian or Langevin dynamics for the CG beads. These internal states help to self-consistently quantify a more complicated set of possible interactions within and between the CG sites, while still maintaining a high degree of coarse-graining in the modeling. The presence of the CG site internal states also greatly expands the possible range of systems amenable to accurate CG modeling, including quite heterogeneous systems such as complex self-assembly processes such as occur for large multi-protein complexes. The development of bottom-up CG models from the underlying atomistic interactions also addresses special challenges in terms of the treatment of solvation, multi-body correlations, representability, transferability, and the missing entropy in CG models. Recent breakthroughs in addressing these issues – in particular by employing developments in machine learning – will be a focus of my talk. As time allows, one "pay-off" application from our multi-year effort will focus on processes from HIV-1 virus replication and especially on the maturation of the HIV-1 virus from several thousand proteins – a phenomenon involving a billion atoms or more over long timescales that cannot be approached through AA MD simulation.
09:30 – 10:00 Cecilia Clementi (Free University Berlin)
Modelling Protein Dynamics with Machine Learning and Molecular Simulation
The last years have seen an immense increase in high-throughput and high-resolution technologies for experimental observation as well as high-performance techniques to simulate molecular systems at a microscopic level, resulting in vast and ever-increasing amounts of high-dimensional data. However, experiments provide only a partial view of macromolecular processes and are limited in their temporal and spatial resolution. On the other hand, atomistic simulations are still not able to sample the conformation space of large complexes, thus leaving significant gaps in our ability to study molecular processes at a biologically relevant scale. We present our efforts to bridge these gaps, by exploiting the available data and using state-of-the-art machine-learning methods to design multiscale models for complex macromolecular systems. We show that it is possible to define simplified molecular models to reproduce the essential information contained both in microscopic simulation and experimental measurements.
10:00 – 10:30 Shoji Takada (Kyoto University)
Protein-resolution modeling for cellular-scale simulations
10:30 – 10:50 Coffee Break
Session 5: Multiscale Methods & Simulation Infrastructure (Chair: TBA)
10:50 – 11:20 Yuji Sugita (RIKEN / University of Tokyo)
TBA
11:20 – 11:50 Bernard R. Brooks (NIH)
An overview of recent applications of machine learning and AI tools for problems in biophysics at the NIH
This presentation consists of a survey of our recent efforts to use machine learning and AI tools for applications in biophysics. Some of the projects presented will include: Leveraging induced polarization for efficient IC50 prediction using minimal features. Predicting iron-sulfur cluster redox potentials using a model derived from protein Structures. Extending protein pKa prediction by tree-based machine learning using PKAD-R, a curated, redesigned and expanded database of experimental pKa values in proteins. Using AI to identify drug repurposing candidates from existing approved drugs that can be used in novel ways for treating unrelated specific cancers.
11:50 – 12:20 Abhishek Singharoy (Arizona State University)
TBA
12:20 – 12:50 Modesto Orozco (IRB Barcelona)
Nucleic Acids simulations from the electron to the chromosome
12:50 – 13:20 Yasuhiro Matsunaga (Saitama University / RIKEN)
TBA
13:20 – 14:30 Lunch
Session 6: From Quantum to ML (Chair: Yuji Sugita)
14:30 – 15:00 Darrin York (Rutgers University)
Multiscale quantum and machine-learning models for drug discovery and enzyme design
This talk presents AI-driven multiscale modeling tools for drug discovery and enzyme design. We introduce end-state alchemical reservoirs and generalized restraint methods for absolute binding free energy calculations, and a Quantum Deep-learning Potential Interaction (QDπ) force field for drug-like molecules that can be applied for bespoke force field generation and indirect “book-ending” alchemical simulations. We further describe a computational enzymology framework that integrates enhanced sampling, free energy methods, and hybrid quantum/ML potentials to investigate enzymatic catalysis. A key development is the Flexible Inner Region Ensemble Separator (FIRES), a multi-component, multi-site approach for maintaining spatial separation between QM solute, QM solvent, and MM solvent regions. These tools enable detailed simulations of catalytic pathways and reveal design principles in natural and engineered nucleic acid enzymes.
15:00 – 15:30 Kwangho Nam (University of Texas at Arlington)
Accelerating Enzyme Catalysis Simulations with Multiscale and Machine Learning-Assisted QM/MM Methods
Understanding enzyme catalysis requires quantum mechanical/molecular mechanical (QM/MM) simulations that are both accurate and capable of accessing biologically relevant time and length scales. However, the high computational cost of conventional QM/MM simulations has long limited their routine application to complex enzymes. In this talk, I will present developments from our group that address these challenges through multiscale and machine learning-assisted QM/MM methodologies. Central to this effort is the multiple time step ai-QM/MM framework, which enable substantially larger time steps while preserving accurate free energy calculations. I will also discuss delta-machine-learning potential approaches that further accelerate QM/MM simulations. Applications to hydride transfer and phosphoryl transfer reactions will be presented, demonstrating how these approaches can efficiently and robustly simulate enzyme catalysis without sacrificing mechanistic detail.
15:30 – 16:00 Tohru Terada (University of Tokyo)
Multiscale Molecular Simulations for Elucidating Protein Function and Assembly Mechanisms
Proteins exhibit diverse functions, spanning from the catalysis of small biomolecule reactions, which necessitates the explicit consideration of electronic structures, to cellular scaffolding, which involves the large-scale assembly of numerous protein molecules. To fully elucidate these diverse protein function mechanisms, multiscale simulation approaches are indispensable. Our research employs hybrid quantum mechanical/molecular mechanical (QM/MM) methods to clarify the catalytic mechanisms of enzymes, and coarse-grained molecular dynamics (CGMD) simulations to investigate the assembly mechanisms of proteins involved in liquid-liquid phase separation (LLPS) and the spontaneous formation of cytoskeletons. In this workshop, I will present our recent findings utilizing QM/MM and CGMD simulations. I will also introduce our latest methodological developments, which include an artificial intelligence (AI)-accelerated transition state search tool and a rigid-body CGMD method specifically designed for large-scale protein assemblies.
16:00 – 16:30 Adrian Roitberg (University of Florida)
Machine Learning Potentials: much faster than quantum, somewhat slower than classical
16:30 – 17:00 Gianni De Fabritiis (University Pompeu Fabra)
Transferable neural network potentials of protein thermodynamics
All-atom molecular simulations offer detailed insights into macromolecular phenomena, but their substantial computational cost hinders the exploration of complex biological processes. We introduce Advanced Machine-learning Atomic Representation Omni-force-field (AMARO), a new neural network potential (NNP) that combines an O(3)-equivariant message-passing neural network architecture, TensorNet, with a coarse-graining map that excludes hydrogen atoms. AMARO demonstrates the feasibility of training coarser NNP, without prior energy terms, to run stable protein dynamics with scalability and generalization capabilities.
17:00 – 17:30 Coffee Break / Poster Preparation

Day 3: May 30, 2026 (Sat)

Session 7: Integrative & Data-Driven Modeling (Chair: Gregory Voth)
09:00 – 09:30 Gerhard Hummer (Max Planck Institute of Biophysics)
Learning from molecular simulations
09:30 – 10:00 Florence Tama (RIKEN / Nagoya University)
TBA
10:00 – 10:30 Alberto Perez (University of Florida)
TBA
10:30 – 10:50 Coffee Break
Session 8: Complex Biomolecular Assemblies & Phase Behavior (Chair: TBA)
10:50 – 11:20 Joan-Emma Shea (UC Santa Barbara)
TBA
11:20 – 11:50 Xiangze Zeng (Hong Kong Baptist University)
Developing Temperature-Dependent Coarse-Grained Potentials for Simulating Phase Separation of Disordered Proteins
Thermoresponsive phase transitions of intrinsically disordered proteins (IDPs), including both upper critical solution temperature (UCST) and lower critical solution temperature (LCST) transitions, are widely observed in natural and synthetic sequences. However, most existing coarse-grained (CG) models with implicit solvents employ temperature-independent interactions and fail to capture solvation-driven LCST behavior. Here, we bridge the gap between atomistic hydrations and macroscopic phase separation. Leveraging extensive all-atom molecular dynamics simulations, we reveal a fundamental linear correlation between the temperature-dependent inter-residue interaction strengths and hydration free energies. Furthermore, we demonstrate that the heterotypic interactions at the molecular level can be well approximated by a simple combination of the homotypic interactions. We incorporate these thermodynamic insights into a physics-based framework, TEA (Temperature-dependent Energetics derived from hydrAtion), which introduces temperature-dependent potentials with minimal phenomenological fitting. The TEA-augmented CG models robustly distinguish UCST- and LCST-type sequences, successfully identify experimentally reported outliers, and accurately reproduce LCST-type single-chain compaction trends and phase diagrams of multiple disordered proteins. Our work provides a transferable and physically interpretable framework that bridges atomistic hydration thermodynamics and phase behavior of IDPs, enabling the simulation of thermoresponsive sequences with minimal phenomenological fitting.
11:50 – 12:20 Ana Damjanovic (Johns Hopkins University)
From Atoms to Neuronal Spikes: A Multiscale Simulation Framework
12:20 – 12:25 Closing Remarks — Gregory Voth (University of Chicago)

Organizing Committee

Gregory Voth (University of Chicago)
Yuji Sugita (RIKEN / University of Tokyo)
Florence Tama (RIKEN / Nagoya University)
Kei Moritsugu (Osaka Metropolitan University)
Yasuhiro Matsunaga (Saitama University / RIKEN)

Co-hosted & Sponsors

Contact

For inquiries, please contact:

Aya Takeuchi - aya.takeuchi at riken.jp

Updates