Simona Biosystems

The N-body problem, which involves predicting the trajectories of bodies by solving Newton's equations of motion, is traditionally limited by the need for small time steps to maintain precision. This requirement can make simulations cumbersome and less practical. To address this, we propose training neural networks to predict these trajectories over larger time steps while preserving accuracy, offering a significant advancement in computational efficiency.

Moreover, the N-body problem presents an ideal benchmark for geometric deep learning, but it has not been fully formalized as such in the literature. In our latest work, we aim to redefine how the N-body problem is used in this context. We employ the Ponita model, a state-of-the-art equivariant graph neural network, to predict the evolution of the system, ensuring physical symmetries are preserved.

To validate the performance of our approach, we introduce a novel suite of "macro-property tests" for a more holistic evaluation of the generated trajectories. This research not only enhances our understanding of the N-body problem but also pushes the boundaries of geometric deep learning. Video bellow shows comparison of trajectories predicted by our neural network and actual trajectories. You can read more about this in this paper.

At Simona Biosystems, we’re pushing the boundaries of molecular dynamics by tackling larger, more complex systems. One such challenge is simulating the solvation of ethane in water, a fundamental process in chemistry and biochemistry. Traditional molecular dynamics simulations often struggle with such systems due to the sheer scale and the intricacies of accurately modeling interactions at the atomic level. These simulations require immense computational resources and can be limited by the need for fine-tuned time steps to preserve precision.

To overcome these challenges, we employ advanced neural networks, including equivariant models, to predict the solvation behavior of ethane in water. By training our models to handle these systems more efficiently, we can use larger time steps without sacrificing accuracy, significantly speeding up the simulation process while maintaining the level of detail needed for accurate predictions.

This approach not only advances our understanding of solvation phenomena but also offers a new perspective for simulating large-scale molecular systems. The ability to accurately model complex solvent-solute interactions opens doors to innovations in drug design, material science, and beyond. Our work in this area demonstrates the power of combining molecular dynamics with state-of-the-art AI, enabling us to explore the atomic world in ways that were previously unimaginable. Video bellow shows trajectories of water and ethane in periodic boundary box predicted by our neural network.