StAIR: Stepwise All-Atom Iterative Reconstruction of Large Coarse-Grained Protein Structures

Overview

Molecular dynamics (MD) simulations model how proteins change conformation during key processes like folding and substrate binding. However, simulating every atom, as is done in all-atom simulations, is computationally expensive for larger proteins. To address this, coarse-graining (CG) groups atoms into fewer “beads”, reducing the number of degrees of freedom in the CG simulation. Still, backmapping, which restores atomic details to these simplified, low resolution CG structures after simulation, remains challenging. Current methods can only accurately backmap from CG structures where one bead represents an amino acid, limiting the speedup of CG simulations. This study introduces StAIR, a novel backmapping framework that successfully backmap from much lower-resolution CG structures using as few as five beads. Unlike the current one-step method, CGVAE, StAIR backmaps twice, employing two machine learning models: the first backmaps to a higher CG resolution, and the second restores full atomic detail. StAIR was trained on real all-atom MD simulation data from the eIF4E and SARS-CoV-2 PLpro proteins and evaluated against CGVAE using root mean square deviation (RMSD) for accuracy and Ramachandran plots for reconstruction quality, comparing backmapped and real structures for each. Overall, StAIR achieved 3x lower RMSD while backmapping from five beads, 10x fewer than what CGVAE can accurately handle, enabling simulations to be run up to 176x faster. Adding a third step further improved reconstruction quality, suggesting how StAIR can be scaled in a similar manner to backmap more complex proteins. StAIR enables highly efficient, biologically accurate simulations, making large-scale protein studies and structure-based drug design significantly faster and more accessible.

Competition history

ISEF 2025 Computational Biology and Bioinformatics · Entry CBIO025

Resources

ISEF abstract page

Related projects

ISEF · 2021

FastFold: Streamlining End-to-End Deep Learning Protein Domain Prediction on COVID-19 Mutations and Other Universal Applications

ISEF · 2026

Novel Nudged Elastic Band (NEB) Approach for Minimum Energy Pathway Modeling in Protein Binding for Efficient Drug Discovery

ISEF · 2024

Elucidating RSV-Host Protein-Protein Interactions: Novel implementation of Graph Attention Networks (GAT) and Structural Visualization

ISEF · 2022

PLCAI: Systematic Pipeline for Protein-Ligand Binding Prediction via Multi-Scale Convolution Neural Network With Applications in Interpretable Drug Design Process

Closest projects by meaning, across every fair and year in the corpus.