Intelligent Resolution: Integrating Cryo-EM with AI-driven Multi-resolution Simulations to Observe the SARS-CoV-2 Replication-Transcription Machinery in Action

Anda Trifan, Defne Gorgun, Zongyi Li, Alexander Brace, Maxim Zvyagin, Heng Ma, Austin Clyde, David Clark, Michael Salim, David J. Hardy, Tom Burnley, Lei Huang, John D. McCalpin, Murali Emani, Hyunseung Yoo, Junqi Yin, Aristeidis Tsaris, Vishal Subbiah, Tanveer Raza, Jessica Liu, Noah Trebesch, Geoffrey Wells, Venkatesh Mysore, Thomas Gibbs, J. C. Phillips, S. Chakra Chennubhotla, Ian Foster, Rick Stevens, Anima Anandkumar, Venkatram Vishwanath, John E. Stone, Emad Tajkhorshid, Sarah A. Harris, Arvind RamanathanbioRxiv 2021 · DOI

TL;DR

This paper introduces a novel AI-driven, multi-resolution simulation framework that integrates cryo-electron microscopy (cryo-EM) data with all-atom molecular dynamics (AAMD) and mesoscale fluctuating finite element analysis (FFEA) simulations. This approach allows researchers to observe the dynamic behavior of complex molecular machines like the SARS-CoV-2 replication-transcription complex (RTC) at unprecedented detail, bridging the resolution and timescale gaps of traditional methods.

What Problem Does This Paper Address?

Understanding how viruses replicate and transcribe their genetic material is crucial for developing effective antiviral therapies. The SARS-CoV-2 replication-transcription complex (RTC) is a key player in this process, but its sheer size and dynamic nature make it incredibly challenging to study. Traditional methods like cryo-EM provide structural snapshots, but often at resolutions that miss crucial dynamic details. All-atom molecular dynamics (AAMD) simulations can capture atomic-level movements, but are computationally expensive and limited to very short timescales, making it impossible to observe the full range of motion for such large complexes. Mesoscale simulations offer a middle ground but often lack the atomic detail needed for precise mechanistic insights. The challenge, therefore, is to bridge these resolution and timescale gaps to gain a comprehensive understanding of how these complex molecular machines function.

Methods

The core innovation is a multi-resolution simulation framework that leverages AI to connect different simulation scales:

  • Cryo-EM Data Integration: Experimental cryo-EM density maps serve as the starting point, providing a low-resolution structural blueprint.
  • Fluctuating Finite Element Analysis (FFEA): This mesoscale method treats biomolecules as visco-elastic continuum solids. It uses a tetrahedral mesh to represent the structure, allowing for simulations at longer timescales and larger length scales than AAMD. FFEA is particularly useful for capturing global conformational changes.
  • All-Atom Molecular Dynamics (AAMD): High-resolution simulations using NAMD are employed to capture atomic-level details and interactions. These are used to refine FFEA simulations and provide detailed potentials for protein-protein interfaces.
  • AI-Driven Integration: A hierarchy of AI methods acts as the glue between FFEA and AAMD.
  • ANCA-AE (Anharmonic Conformational Analysis Autoencoder): This method learns to summarize global conformational changes from both FFEA and AAMD simulations, identifying key states and transitional points.
  • CVAE (Convolutional Variational Autoencoder): Used to describe localized conformational changes within specific protein subunits.
  • Graph Neural Operators (GNO): A novel approach to capture time-dependent conformational changes from AAMD simulations, enabling prediction of future states and bridging the gap between simulation modalities.
  • Multi-Site Workflow Management: The Balsam workflow manager orchestrates these diverse computational tasks across multiple supercomputing facilities, optimizing resource utilization and enabling coordinated execution of simulations and AI analyses.

Datasets

The primary dataset used is the cryo-EM density map of the SARS-CoV-2 RTC (EMD-31138) and corresponding atomic structures (e.g., 7EGQ.pdb). The AAMD simulations generate extensive trajectory data that serves as input for the AI models. No new benchmarks are explicitly introduced, but the paper demonstrates the framework's capability on the SARS-CoV-2 RTC system.

Key Insights and Ideas

  • Bridging the Resolution Gap: The framework successfully integrates low-resolution cryo-EM data with high-resolution AAMD simulations, providing a more complete picture of molecular dynamics.
  • AI as a Unifying Force: AI models are crucial for learning the relationships between different simulation scales, inferring interface potentials, and guiding the overall simulation process.
  • Observing Complex Dynamics: The approach allows for the visualization of intricate conformational changes within the RTC, such as RNA unwinding and the coordinated motions of its subunits.
  • Enabling Drug Discovery: By providing unprecedented access to the RTC's dynamics, this work can inform the design of new antiviral drugs that target specific interactions within the complex.
  • Scalable and Distributed Science: The use of Balsam demonstrates a robust method for coordinating complex computational workflows across multiple HPC centers, a critical capability for modern scientific research.

Key Results

  • The framework successfully models the SARS-CoV-2 RTC, capturing conformational fluctuations and dynamics that were previously inaccessible.
  • The AI methods effectively summarize complex simulation data, revealing key conformational states and transitions.
  • The multi-site workflow demonstrated near 100% utilization on HPC resources, significantly accelerating the overall time to solution.
  • The study provides insights into the coordinated motions of RTC subunits, particularly highlighting the role of interface proteins like nsp8, nsp9, and nsp10.
  • The paper showcases the potential for AI to predict time-dependent conformational changes, reducing the need for costly integration steps.

Limitations and Future Work

  • Computational Cost: While the multi-resolution approach significantly reduces computational burden compared to pure AAMD, it still requires substantial HPC resources.
  • AI Model Optimization: Some AI models (e.g., GNO) were not fully optimized for performance, suggesting potential for further speedups.
  • Data Transfer Bottlenecks: Coordinating large file transfers across supercomputing sites for visualization and analysis remains a challenge.
  • Future Work: The authors suggest extending this AI-enabled multi-resolution simulation framework to other large biomolecular machines and exploring its application in drug discovery and other complex scientific domains. The development of more efficient AI models and improved data management strategies are also implied future directions.

Summary updated

Abstract

ABSTRACT The severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) replication transcription complex (RTC) is a multi-domain protein responsible for replicating and transcribing the viral mRNA inside a human cell. Attacking RTC function with pharmaceutical compounds is a pathway to treating COVID-19. Conventional tools, e.g., cryo-electron microscopy and all-atom molecular dynamics (AAMD), do not provide sufficiently high resolution or timescale to capture important dynamics of this molecular machine. Consequently, we develop an innovative workflow that bridges the gap between these resolutions, using mesoscale fluctuating finite element analysis (FFEA) continuum simulations and a hierarchy of AI-methods that continually learn and infer features for maintaining consistency between AAMD and FFEA simulations. We leverage a multi-site distributed workflow manager to orchestrate AI, FFEA, and AAMD jobs, providing optimal resource utilization across HPC centers. Our study provides unprecedented access to study the SARS-CoV-2 RTC machinery, while providing general capability for AI-enabled multi-resolution simulations at scale.
BibTeX
@misc{trifan2021intelligent,
  author = {Anda Trifan and Defne Gorgun and Zongyi Li and Alexander Brace and Maxim Zvyagin and Heng Ma and Austin Clyde and David Clark and Michael Salim and David J. Hardy and Tom Burnley and Lei Huang and John D. McCalpin and Murali Emani and Hyunseung Yoo and Junqi Yin and Aristeidis Tsaris and Vishal Subbiah and Tanveer Raza and Jessica Liu and Noah Trebesch and Geoffrey Wells and Venkatesh Mysore and Thomas Gibbs and J. C. Phillips and S. Chakra Chennubhotla and Ian Foster and Rick Stevens and Anima Anandkumar and Venkatram Vishwanath and John E. Stone and Emad Tajkhorshid and Sarah A. Harris and Arvind Ramanathan},
  title = {{Intelligent Resolution: Integrating Cryo-EM with AI-driven Multi-resolution Simulations to Observe the SARS-CoV-2 Replication-Transcription Machinery in Action}},
  year = {2021},
  doi = {10.1101/2021.10.09.463779},
  howpublished = {bioRxiv},
}
Figure