AMARO: All Heavy-Atom Transferable Neural Network Potentials of Protein Thermodynamics

Abstract: All-atom molecular simulations offer detailed insights into macromolecular phenomena, but their substantial computational cost hinders the exploration of complex biological processes. We introduce Advanced Machine-learning Atomic Representation Omni-force-field (AMARO), a new neural network potential (NNP) that combines an O(3)-equivariant message-passing neural network architecture, TensorNet, with a coarse-graining map that excludes hydrogen atoms. AMARO demonstrates the feasibility of training coarser NNP, without prior energy terms, to run stable protein dynamics with scalability and generalization capabilities.

• The paper introduces AMARO, an innovative neural network potential that integrates a coarse-grained heavy-atom mapping with TensorNet to model protein thermodynamics efficiently.
• It employs an O(3)-equivariant message-passing architecture and variational force matching, achieving a low L1 test loss of 5.07 kcal/mol/Å and robust performance on fast-folding proteins.
• AMARO reduces computational cost while accurately sampling free energy landscapes, enabling scalable simulations of larger protein domains for advanced biological research.

AMARO: All Heavy-Atom Transferable Neural Network Potentials of Protein Thermodynamics

The paper entitled "AMARO: All Heavy-Atom Transferable Neural Network Potentials of Protein Thermodynamics" presents an innovative approach to all-atom molecular simulations. The significant computational cost traditionally associated with such methods has been a pervasive obstacle in exploring complex biological processes. The authors introduce AMARO, a neural network potential (NNP) combining an O(3)-equivariant message-passing neural network architecture, TensorNet, with a coarse-graining (CG) map that excludes hydrogen atoms. This paper elucidates the potential AMARO offers in balancing computational efficiency with accurate representation of protein dynamics.

Core Methodology and Components

The AMARO methodology collates several advanced techniques:

1. CG Mapping Without Hydrogen Atoms: The authors adopt a CG map that removes hydrogen atoms, focusing instead on heavy atoms. This approach leverages the physical intuition that hydrogen atoms can be computationally costly with relatively minor contributions to Lennard-Jones interactions.
2. TensorNet Neural Network Architecture: TensorNet, an O(3)-equivariant message-passing neural network, forms the backbone of the model. O(3)-equivariance ensures rotationally and translationally invariant force field representations, crucial for modeling physical systems.
3. Variational Force Matching Technique: The force matching is approached using a mean-squared deviation minimization between CG candidate force fields and atomistic forces. This method establishes a robust framework for learning the potential of mean force (PMF) effectively, even at CG resolutions.

Datasets and Training

The study utilized the extensive mdCATH dataset with significant filtering criteria to refine the training set, resulting in 2,834 domains and over 26 million conformations. The filtered set was divided into training, validation, and testing splits. TensorNet was trained for 100 epochs, achieving notable accuracy (L1 test loss of 5.07 kcal/mol/Å).

Validation and Generalization

The generalization capability of AMARO was rigorously tested on larger protein domains and unseen fast-folding proteins:

1. Scale-Up Validation: Testing on larger protein domains with 150–250 residues demonstrated the model’s robust scalability, maintaining a mean absolute error (MAE) of 4.98 kcal/mol/Å.
2. Fast-Folding Proteins: For proteins not in the training set (Chignolin, Trp-Cage, Villin, and α3D), AMARO successfully recovered native structures, providing high equilibrium probabilities and low RMSD values when compared to experimental crystal structures.

Results and Performance

AMARO outperforms traditional all-atom simulations in terms of computational efficiency. Analyzing both accuracy and sampling efficiency underscores the model’s ability to effectively explore free energy landscapes. The efficiency is especially evident in broader conformational space coverage and reduced computational time due to fewer degrees of freedom in CG mapping.

Implications and Future Directions

AMARO’s demonstrated efficiency and accuracy have substantial implications for biological research and drug discovery. By simplifying computational workflows without compromising on dynamical accuracy, AMARO facilitates the study of larger and more complex biological systems within feasible time frames.

Future developments should focus on improving computational efficiency and reducing memory usage. Enhancements may include optimizing TensorNet’s architecture and further refining CG maps to encompass more complex interactions, e.g., solvent interactions, which are currently simplified. Extending the model to incorporate explicit solvation effects and improving the representation of highly interactive groups like NH3+ can further boost AMARO’s utility in various biological contexts.

Conclusion

Overall, AMARO represents a significant step forward in the field of molecular dynamics simulations. By leveraging advanced machine-learning techniques and an innovative CG mapping strategy, the authors present a robust and scalable model for accurate protein thermal dynamics simulation. This work lays the groundwork for future explorations and refinements in NNP frameworks, challenging existing paradigms and offering a viable path towards more efficient and expansive molecular simulations.