AIMD-Chig: Exploring the conformational space of a 166-atom protein Chignolin with ab initio molecular dynamics

Tong Wang, Xinheng He, Mingyu Li, Bin Shao, Tie‐Yan LiuScientific 2023 · DOI

Abstract

Molecular dynamics (MD) simulations have revolutionized the modeling of biomolecular conformations and provided unprecedented insight into molecular interactions. Due to the prohibitive computational overheads of ab initio simulation for large biomolecules, dynamic modeling for proteins is generally constrained on force field with molecular mechanics, which suffers from low accuracy as well as ignores the electronic effects. Here, we report AIMD-Chig, an MD dataset including 2 million conformations of 166-atom protein Chignolin sampled at the density functional theory (DFT) level with 7,763,146 CPU hours. 10,000 conformations were initialized covering the whole conformational space of Chignolin, including folded, unfolded, and metastable states. Ab initio simulations were driven by M06-2X/6-31 G* with a Berendsen thermostat at 340 K. We reported coordinates, energies, and forces for each conformation. AIMD-Chig brings the DFT level conformational space exploration from small organic molecules to real-world proteins. It can serve as the benchmark for developing machine learning potentials for proteins and facilitate the exploration of protein dynamics with ab initio accuracy.

Introduction

Molecular Dynamics (MD) simulations serve as "computational microscopes" for biological systems, yet they face a fundamental trade-off between speed and accuracy. Classical MD relies on empirical force fields—mathematical approximations where atoms are treated as spheres with fixed charges. While computationally efficient, these models often fail to capture critical electronic effects and polarization, leading to inaccuracies in complex processes such as protein folding.

Ab Initio Molecular Dynamics (AIMD) offers a solution by solving Schrödinger’s equation at every step, providing high-fidelity results. However, the computational cost of AIMD has historically restricted its use to small organic molecules. Machine Learning Potentials (MLPs) have emerged as a promising middle ground, aiming to replicate quantum accuracy at classical speeds. The bottleneck for MLPs in protein research has been the lack of diverse, high-quality quantum-level training data. AIMD-Chig addresses this "data desert" by providing a massive dataset for the 166-atom protein Chignolin.

Methods

The researchers developed a multi-stage pipeline to efficiently sample the conformational space of Chignolin without the need for a single, prohibitively long quantum simulation:

  1. Broad Sampling: Classical MD (including Replica Exchange) was used to generate an extensive library of conformations, encompassing folded, unfolded, and intermediate states.
  2. Dimensionality Reduction: Time-lagged Independent Component (tIC) analysis projected these complex shapes into a lower-dimensional energy landscape to identify key structural variations.
  3. Anchor Selection: 10,000 "anchor" structures were selected to represent diverse points across the energy landscape.
  4. Quantum Refinement: Short (225 fs) AIMD simulations were performed for each anchor at the DFT level (M06-2X/6-31G*).
  5. Data Extraction: After discarding initial equilibration frames, 200 high-quality snapshots per anchor were retained, totaling 2 million samples.
Overview of the AIMD-Chig data generation pipeline, from initial structures to quantum MD simulation generation.

The AIMD-Chig Dataset

The primary contribution of this work is the AIMD-Chig dataset, which represents a significant leap in scale for quantum protein modeling. The dataset includes coordinates, energies, and atomic forces for 2 million conformations, generated at a total cost of over 7.7 million CPU hours. Unlike previous benchmarks like MD17 or MD22, AIMD-Chig provides the density and structural diversity necessary to train models capable of understanding a protein's global folding landscape.

Organization and structure of the AIMD-Chig dataset, including validation protocols and data splits.

Key Insights and Results

  • The Importance of Metastable States: The study found that 10,000 anchors were required to accurately map the transition shapes between folded and unfolded states. Fewer anchors resulted in a fragmented view of the energy landscape.
  • Quantifying Classical Error: The researchers quantified the limitations of classical molecular mechanics (MM), which showed an energy error of approximately 21.7 kcal/mol compared to DFT. Semi-empirical methods also exhibited significant deviations, reinforcing the necessity of full DFT data.
  • Achieving Chemical Accuracy: Training a state-of-the-art Graph Neural Network (VisNet) on the full dataset reduced the Mean Absolute Error (MAE) to 0.738 kcal/mol for energy and 0.195 kcal/mol/Å for forces, reaching the threshold of "chemical accuracy."
  • Simulation Stability: ML potentials trained on AIMD-Chig demonstrated the ability to run stable 10-nanosecond simulations while maintaining the protein's structural integrity, a common challenge for ML models trained on sparse data.

Conclusion and Future Work

While Chignolin is a relatively small protein at 166 atoms, AIMD-Chig demonstrates that quantum-level accuracy is achievable for real-world biological systems. The computational investment required highlights the challenges of scaling this approach to larger enzymes or receptors. Future developments will likely focus on incorporating explicit solvent environments to better capture the complexity of biological systems, requiring further innovations in sampling strategies and computational efficiency.

Summary updated

Abstract

Molecular dynamics (MD) simulations have revolutionized the modeling of biomolecular conformations and provided unprecedented insight into molecular interactions. Due to the prohibitive computational overheads of ab initio simulation for large biomolecules, dynamic modeling for proteins is generally constrained on force field with molecular mechanics, which suffers from low accuracy as well as ignores the electronic effects. Here, we report AIMD-Chig, an MD dataset including 2 million conformations of 166-atom protein Chignolin sampled at the density functional theory (DFT) level with 7,763,146 CPU hours. 10,000 conformations were initialized covering the whole conformational space of Chignolin, including folded, unfolded, and metastable states. Ab initio simulations were driven by M06-2X/6-31 G* with a Berendsen thermostat at 340 K. We reported coordinates, energies, and forces for each conformation. AIMD-Chig brings the DFT level conformational space exploration from small organic molecules to real-world proteins. It can serve as the benchmark for developing machine learning potentials for proteins and facilitate the exploration of protein dynamics with ab initio accuracy.
BibTeX
@inproceedings{wang2023aimd,
  author = {Tong Wang and Xinheng He and Mingyu Li and Bin Shao and Tie‐Yan Liu},
  title = {{AIMD-Chig: Exploring the conformational space of a 166-atom protein Chignolin with ab initio molecular dynamics}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41597-023-02465-9},
}
Figure