Minisymposium

To fully understand biological functions, high-resolution biomolecular structures are required, which can be obtained through diverse experimental methods such as X-ray crystallography and, more recently, cryo-electron microscopy (cryo-EM) and computationally via AlphaFold2. While these techniques provide valuable structural insights, elucidating the dynamic of these biomolecules is also essential, given that some molecules display inherent structural flexibility. A thorough comprehension of such dynamics involves characterizing various conformational states at the atomic level. Raw experimental data from cryo-EM consist of millions of 2D images that represent specific views of a biomolecule, potentially capturing different conformational states, as each image may correspond to a distinct conformation. To connect these 2D images with atomic models, Molecular Dynamics (MD) simulations can be employed to generate conformations that align with the observed data. Due to the large scale of cryo-EM datasets, processing them often demands high-performance computing. In this talk, I will present our work and the implementation of MDSPACE, a biased MD simulation method, for conducting such analyses on the Fukagu system.

Molecular dynamics (MD) simulations generate extensive data to analyze molecular behaviors, necessitating supercomputers for execution and post-simulation analysis. Traditional approaches store simulation data for later analysis, causing bottlenecks that delay discoveries and limit the analysis of saved data. Performing data processing and analysis as generated eliminates storage overheads and reduces I/O costs and time. This talk examines a software framework for early termination of MD simulations using collective variables (CVs) at runtime. Early termination can save computational resources by redirecting them to unexplored conformational space regions. CVs identify significant conformational states, such as a folded protein structure, and trigger early termination. We compare full simulations of the FS peptide on the Summit supercomputer with early-terminated simulations using CVs like Largest Eigen Value (LEV) and Effective Sample Size (ESS). Results show that LEV and ESS terminated simulations preserve critical conformational states. Root Mean Square Deviation (RMSD) analysis confirms the alignment of conformational spaces, and Hidden Markov Models (HMM) demonstrate maintained dynamic behavior. Our framework's early termination offers a robust representation of conformational space while optimizing computational efficiency.

Multiscale molecular dynamics workflows, such as MuMMI, integrate coarse-grain and all-atom simulations to model complex biomolecular systems, but they introduce substantial data movement and synchronization challenges. In this talk, we present Analytics4X (A4X), a modular benchmarking framework that models producer-consumer patterns using state machine abstractions. Focusing on one-to-one data movement, we evaluate performance on LLNL’s Corona supercomputer across storage systems, including Lustre and DYAD. Results show that DYAD’s adaptive synchronization and efficient communication can deliver a speedup of up to 194×, offering key insights for building scalable, ML-driven MD workflows with tightly coupled simulation and analysis.

In this talk, we explore the concept of hybrid classical-quantum systems as a response to the growing demand for computational resources. Modern scientific applications increasingly rely on diverse hardware accelerators like GPUs, TPUs, and NPUs. With the advent of the post-Moore era and the rise of quantum computing, new opportunities and challenges emerge as we aim to integrate quantum systems into the established computational continuum. Hybrid systems promise unprecedented resource efficiency but require addressing key challenges such as data encoding, noise, transpilation, and the heterogeneity of quantum architectures - necessitating new hardware and software layers.In the second part of the talk we discuss the challenges and design choices for the hybrid classic quantum system by utilizing the molecular dynamics applications as an use case.