Predicting Molecular Trajectories using Machine Learning Methods

Predicting molecular trajectories is a cornerstone of computational chemistry, with implications for drug discovery and molecular dynamics simulations. This study presents a comprehensive analysis of various machine learning models for the prediction of aspirin molecular trajectories, as captured in a dataset of 1500 frames calculated via the CCSD [Psi4, cc-pVDZ] method. We explore statistical sampling methods, including random walk and Monte Carlo Markov Chain (MCMC), alongside a suite of neural networks comprising feed-forward, recurrent neural network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) architectures. Additionally, we investigate the potential of Equivariant Neural Networks (E3NN) to enforce permutation and rototranslational invariance, as well as Graph Convolutional Networks (GCN) for leveraging the inherent graph structure of molecules. Our results highlight the comparative effectiveness of these methods, with GCNs unexpectedly outperforming others in trajectory prediction accuracy. The study also delves into the novel application of diffusion models, treating molecular pose prediction as a generative problem, despite challenges in maintaining physical plausibility. Though preliminary, our findings underscore the promise of graph-based methods in capturing molecular interactions and dynamics, paving the way for future advancements in efficient and accurate trajectory prediction in computational chemistry.