The current state of modeling protein motions : from physics to AI
October 7 @ 11:00 – 12:00 CEST
Invited speaker:
Sergei Grudinin, LJK CNRS Grenoble, France
Artificial intelligence, and more specifically deep learning, has recently emerged as a powerful approach to exploit the massive amounts of protein sequence and structure data available nowadays toward guiding biological intervention to improve human health. A couple of months ago, the alphaFold2 architecture from DeepMind revolutionised the field of protein structure prediction by reaching unprecedented levels of near-experimental accuracy. This achievement has been made possible mostly thanks to the latest improvements in geometric learning and natural language processing (NLP) techniques.
While the problem of determining how a protein folds in three dimensions (3D) is essentially solved, accessing protein motions is becoming more central than ever before [1]. At the European level, the ELIXIR community is investing efforts right now to create a comprehensive resource for structural diversity and flexibility in the Protein Data Bank (PDB), which contains all experimentally-determined protein 3D structures. Indeed, proteins are flexible biological objects, constantly moving and changing their shape to interact with their environment and cellular partners. This inherent flexibility is highly relevant for protein functioning. Experimentally, it is very difficult to observe proteins directly in action, and we have mostly access to isolated clusters of “snapshots” (conformations) representative of a few functional states.
I will present relatively simple physics-based models developed in our team, where the protein is represented by an elastic network. They have proven very useful to nonlinearly extrapolate functional motions, starting from a single structure and predict structural protein transitions [2-4]. I will also show an extension of these developments to construct a multi-level representation of protein flexibility. Then, I will outline the current state of AI methods to model protein structural heterogeneity and connect it with the physics-based models.
References
[1] Laine, Elodie, et al. “Protein sequence-to-structure learning: Is this the end (-to-end revolution)?.” Proteins in Press (2021).
[2] Laine, Elodie, and Sergei Grudinin. “HOPMA: Boosting protein functional dynamics with colored contact maps.” The Journal of Physical Chemistry B 125.10 (2021): 2577-2588.
[3] Grudinin, Sergei, Elodie Laine, and Alexandre Hoffmann. “Predicting protein functional motions: an old recipe with a new twist.” Biophysical journal 118.10 (2020): 2513-2525.
[4] Hoffmann, Alexandre, and Sergei Grudinin. “NOLB: Nonlinear rigid block normal-mode analysis method.” Journal of chemical theory and computation 13.5 (2017): 2123-2134.
Google Calendar 
