Oral Presentation 23rd International Society of Magnetic Resonance Conference 2023

Expedited and facilitated NMR analysis of proteins for difficult cases assisted by MD simulation and machine learning. (#139)

Naohiro Kobayashi 1 , Tsuyoshi Konuma 2 , Kyohei Arita 2 , Shintaro Minami 3 , Kohya Sakuma 4 , Nobuyasu Koga 3 4 5
  1. RIKEN Center for Biosystems Dynamics Research (RIKEN BDR), Yokohama, KANAGAWA, Japan
  2. Graduate School of Medical Life Science, Yokohama City University, Yokohama, Kanagawa, Japan
  3. Institute for Molecular Science (IMS), National Institutes of Natural Sciences (NINS), Okazaki, Aich, Japan
  4. Department of Structural Molecular Science, School of Physical Sciences, SOKENDAI, Hayama, Kanagawa, Japan
  5. xploratory Research Center on Life and Living Systems (ExCELLS), National Institutes of Natural Sciences, Okazaki, Aich, Japan

Conventional solution NMR analysis of small proteins provides information not only on the 3D structure model and protein-protein interactions, but also on the dynamic states and structural exchanges upon local conformational changes and complex formation, which would be helpful for understanding the mechanistic information related to their biological functions. There are many cases that cannot be solved by X-ray crystallography or cryo-electron microscopy, especially in the case of small proteins with a low tendency to crystallization or low binding affinity (Kd 10 - 100 uM), mainly because the targeted protein or the complex system has some regions with high flexibility. On the other hand, in analysis of the solution NMR samples having a poor crystallization, we often encounter obstacles, such as overlapping, line broadening, or missing NMR signals which make assignments difficult. In this study, we report successful cases of studying de novo designed helix-rich proteins [1,2], and two small protein complexes in which Deep Neural Networks (DNNs) and MD calculations [3,4] are effectively utilized to expedite complex analysis. A core technique that was used for fully automated NMR signal analysis is assisted by fully automated NMR signal detection and noise filtration using DNNs [5]. One of the major key points of our strategy is the backbone chemical shifts predicted from modeled NMR structure ensemble using DNNs for 15N, 13Ca, 13Cb, and 13CO atoms. Another key point is surveying dynamic states of proteins and complex formation on the usec to msec time scales as exploring potential mean force (PMF) derived from MD simulations based on extended ensemble methods. We have designed a new strategy combining these key technologies to facilitate NMR, especially for more difficult cases, which will be presented in the conference.

[1] Sakuma, K. et al., https://www.biorxiv.org/content/10.1101/2021.07.14.449347v1
[2] Minami, S. et al., https://www.biorxiv.org/content/10.1101/2021.08.06.455475v1
[3] Hatta, K. et al., Nuc. Acid. Res. (2022) 50(21), 12527-12542.
[4] paper in preparation.
[5] Kobayashi, N. et. al, Bioinformatics (2018) 34(24), 4300-4301.