A Neural Time-Series Learning Method for Accelerating Free-Energy Perturbation and Rare-Event Molecular Dynamics...

Area of Science:

• Computational chemistry
• Materials science
• Drug discovery

Background:

• Molecular dynamics (MD) simulations are crucial for materials and drug discovery but are computationally intensive.
• Existing sequence-based accelerators struggle with long-range temporal structures and noisy data in MD trajectories.

Purpose of the Study:

• To introduce BiLSTMK-MD, a novel neural time-series learning method to create a surrogate for MD and FEP trajectories.
• To reduce the sampling requirements for free-energy estimation and rare-event characterization.

Main Methods:

• BiLSTMK-MD couples a bidirectional LSTM encoder with an attention mechanism and a Kolmogorov-Arnold network output layer.
• A two-stage, fANOVA-guided Bayesian optimization tunes hyperparameters for optimal performance.
• The model constructs a causality-preserving surrogate for MD and FEP trajectories.

Main Results:

• Achieved mean absolute errors below 1.5 kcal mol^-1 for free-energy increments.
• Reconstructed dihedral free-energy basins using only 1-10% of simulation trajectories.
• Demonstrated up to 400-fold acceleration for FEP and >700-fold speedup for rare-conformation sampling.

Conclusions:

• BiLSTMK-MD offers a significant acceleration for MD simulations, particularly for FEP and rare-event sampling.
• This neural time-series surrogate provides an efficient route to reduce computational demands in molecular simulations.