Abstract
Biomolecular dynamics simulation is a fundamental technology for life sciences research, and its usefulness depends on its accuracy and efficiency. Classical molecular dynamics simulation is fast but lacks chemical accuracy. Quantum chemistry methods like density functional theory (DFT) can reach chemical accuracy but cannot scale to support large biomolecules. We introduce an AI-based ab initio biomolecular dynamics system (AI²BMD) that can efficiently simulate large biomolecules with ab initio accuracy. AI²BMD uses a protein fragmentation scheme and machine learning force field to achieve generalizable ab initio accuracy for energy and force calculations for various proteins comprising over 10,000 atoms. Compared to DFT, it reduces computational time by several orders of magnitude. With several hundred nanoseconds of dynamics simulations, AI²BMD demonstrated its capability of efficiently exploring the conformational space of peptides and proteins, deriving accurate $^3J$-couplings that match NMR experiments, and showing protein folding and unfolding tendencies. Furthermore, AI²BMD enables precise free energy calculations for protein folding, and the estimated melting temperatures are well aligned with experiments. AI²BMD could potentially complement wet-lab experiments, detect the dynamic processes of bioactivities, and enable biomedical research that is currently impossible to conduct.
Introduction
In computational biology, researchers have long faced a "speed vs. accuracy" trade-off. Classical Molecular Dynamics (MD) uses simplified "ball-and-spring" models that are fast but lack the chemical accuracy needed to capture complex electronic interactions. Conversely, quantum chemistry methods like Density Functional Theory (DFT) provide high precision but are computationally prohibitive, with costs scaling cubically ($O(N^3)$) with the number of atoms. This makes simulating large proteins for biologically relevant timescales impossible. AI²BMD addresses this bottleneck, enabling the exploration of protein folding and dynamics with quantum-level precision at a fraction of the cost.
Methods: A "Divide and Conquer" Approach
The core innovation of AI²BMD is its fragmentation-based machine learning force field. Instead of calculating the quantum state of a massive protein as a single entity, the system employs a multi-step workflow:
- Fragmentation: Proteins are broken into overlapping "protein units" (dipeptides). Since there are only 21 standard types of these units, the complexity is significantly reduced.
- ViSNet Force Field: A deep learning model, ViSNet, is trained on a massive dataset of 20.88 million dipeptide conformations calculated at the high-fidelity M06-2X/6-31g* quantum level.
- Linear Scaling: By processing fragments and reassembling them, the computational complexity grows linearly ($O(N)$) with the number of atoms.
Results and Validation
AI²BMD was evaluated on nine diverse proteins, ranging from the 175-atom Chignolin to the 13,728-atom Aminopeptidase N.
Accuracy and Speed
The system outperformed conventional molecular mechanics (MM) by two orders of magnitude in energy accuracy, achieving a Mean Absolute Error (MAE) of 0.023 kcal/mol compared to 3.198 kcal/mol for MM. Crucially, AI²BMD achieves a speedup of approximately 10,000x over DFT for a 1,000-atom protein, reducing calculation time from 230 hours to just 44 seconds per step.
Experimental Alignment
Beyond theoretical metrics, AI²BMD matches real-world experimental data:
- NMR Matching: It achieved a Pearson correlation of 0.81 in reproducing $^3J$-couplings, significantly outperforming classical force fields (0.59).
- Thermodynamics: For the WW domain protein, AI²BMD predicted a melting temperature ($T_m$) of 359.06 K, closely aligning with the experimental value of 371 K.
Discussion and Future Directions
AI²BMD represents a significant leap toward generalizable ab initio simulations. Unlike previous models overfit to specific structures, AI²BMD's unit-based training allows it to simulate virtually any protein without retraining.
Future work will focus on further closing the efficiency gap with classical MD through engineering optimizations and batch inference. Additionally, the researchers plan to expand the system to include lipids, DNA/RNA, and complex solvent interfaces, paving the way for more accurate drug discovery and the study of dynamic biological processes.
Summary updated