NCAFM2023 Programme Booklet

Wednesday 1720-1740

SIMULATING ADSORPTION, SELF-ASSEMBLY AND HYDRATION STRUCTURES OF PEPTIDES IN WATER THROUGH SCALABLE NEURAL NETWORK POTENTIALS

F. Priante 1 (Presenting), A. Yurtsever 2 , T. Fukuma 2 and A. S. Foster 1, 2

1 Department of Applied Physics, Aalto University, Helsinki FI-00076, Finland 2 WPI Nano Life Science Institute (WPI-Nano LSI), Kanazawa University, Japan Email: fabio.priante@aalto.fi

The vast combinatorial space accessible within short peptide sequences has allowed to identify specific aminoacid chains with a strong binding affinity for solid surfaces. Studying the self-assembly of these peptides could not only hold promise for nanobioelectronics applications, but it might also provide clues regarding protein aggregation in neurodegenerative diseases. Atomic Force Microscopy (AFM) is an ideal tool for probing such self-assembled peptide nanostructures, as it enables real-time and on-site imaging within a physiological aqueous environment [1]. In the past few years, neural network potentials (NNPs) have gained prominence due to their ability to achieve near-density functional theory (DFT) accuracy with significantly reduced computational requirements [2]. However, their application has usually been restricted to relatively small systems, where the challenges of robustness and scalability of NNPs are less critical [3, 4]. Recently, efforts have been made in facing these issues, leading to efficient parallelization of NNPs across multiple graphics processing units (GPUs) [5]. In this work, we employ scalable equivariant NNPs to model a large and complex system involving the self-assembly of peptides on a graphite substrate in an aqueous solution. We rapidly evaluate hundreds of candidate self-assembled structures, finally comparing the water density profiles of the most stable ones with 3D-Atomic Force Microscopy [6] experiments. By demonstrating the effectiveness of NNPs for simulating large-scale solid-liquid interfaces, our study presents a new avenue for research that previously relied exclusively on classical molecular dynamics methods. References [1] A. Yurtsever, L. Sun, K. Hirata, T. Fukuma, S. Rath, H. Zareie, S. Watanabe, M. Sarikaya, ACS Nano, 2023, 17(8) [2] S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, B. Kozinsky, Nature Communications, 2022, 13(1). [3] S. Stocker, J. Gasteiger, F. Becker, S. Günnemann, J. T. Margraf, Machine Learning: Science and Technology, 2022, 3 (4), 045010. [4] X. Fu, Z. Wu, W. Wang, T. Xie, S. Keten, R. Gomez-Bombarelli, T. Jaakkola, https://arxiv.org/abs/2210.07237 [5] A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, B. Kozinsky, Nature Communications, 2023, 14(1). [6] T. Fukuma, R. Garcia, ACS Nano, 2018, 12(12), 11785–11797.

67