Topological graphs in MD simulations and machine learning LLMs (large language models) for spectroscopies

In this presentation, I will review some of our developments and applications of the past few years in algorithmic graph theory. These graphs have been developed in order to analyze the conformational dynamics arising in molecular dynamics simulations of complex molecular systems (ab initio MD, classical FF-MD). An important aspect is that our topological graphs are transferable without modification from ’simple’ gas phase molecules, to liquids, to more complex inhomogeneous interfaces between solid and liquid for instance, as will be illustrated during the talk.

We have developed topological graphs with two levels of granularity : atomistic 2D-MolGraphs^1-3 and coarse-grained polygraphs of H-Bonded cycles.⁴ These graphs have been implemented with the key-algorithms of isomorphism and polymorphism. The graphs have also been included in computational workflows.^5-6

We will show in our presentation that the use of algorithmic graph theory provides a direct and fast approach to identify the actual conformations that are sampled over time in a trajectory. Graph of transitions can be extracted, showing at one glance all the interconversions over time between these conformations. H-Bonded networks in condensed matter molecular systems, such as for instance at aqueous interfaces, are shown to be easily captured with graphs.^7,2-3 We will also show how the 2D-MolGraphs can easily be included in automated high-throughput in silico reactivity workflows, and how essential they are in some of the decisive steps to be taken in these workflows.^5-6 Recently, we developed coarse grained polygraphs of H-Bonded cycles,⁴ that we will show to be essential graphs for the analysis of the dynamics of flexible molecules such as a peptides and more complex biomolecules. These graphs can be extended to H-Bonded networks in condensed matter systems. Further recent developments include the recognition of structural motifs at solid/liquid water interfaces. With motifs in hand, the detailed structure of liquid water in the Binding Interfacial Layer (BIL) is directly known and can be directly used in order to interpret the SFG (Sum Frequency Generation) spectroscopic signatures at these aqueous interfaces.

We are also developing machine learning techniques with Large Language Models (LLMs) in order to predict spectroscopic signals (IR, Raman, SFG). Our first results will be shown in this presentation.

These works have been achieved with the following collaborators :
1 Dr Sana Bougueroua, 1 Dr Alvaro Cimas, 2 Prof Dominique Barth, 2 Prof Jérémie Cabessa, 2 Ylène Aboulfath, 3 Dr Ali Hashemi, 3 Prof Evgeny Pidko

1Université Paris-Saclay, Univ Evry, CY Cergy Paris Université, CNRS, LAMBE, 91025 Evry-Courcouronnes, France
2Université Paris-Saclay, Univ Versailles Saint Quentin, DAVID, 78035 Versailles, France
3Inorganic Systems Engineering, Department of Chemical Engineering, Faculty of Applied Sciences, Delft University of Technology, 2629 HZ Delft, The Netherlands

References :
1- S. Bougueroua, R. Spezia, S. Pezzotti, S. Vial, F. Quessette, D. Barth, M.-P. Gaigeot
Graph theory for automatic structural recognition in molecular dynamics simulations.
J. Chem. Phys. 149:184102-15 (2018)
2- S. Bougueroua, M. Bricage, Y. Aboulfath, D. Barth, M.-P. Gaigeot
Algorithmic Graph Theory, Reinforcement Learning and Game Theory in MD Simulations : From 3D Structures to Topological 2D-Molecular Graphs (2D-MolGraphs) and Vice Versa
Molecules 28 : 2892-2912 (2023)
DOI : 10.3390/molecules28072892
3- S. Bougueroua, Y. Aboulfath, D. Barth, M.-P. Gaigeot
Algorithmic graph theory for post-processing molecular dynamics trajectories
Mol. Phys. e2162456 (2023)
DOI : 10.1080/00268976.2022.2162456
4- Y. Aboulfath, S. Bougueroua, A. Cimas, V. Chantitch, D. Barth, M.-P. Gaigeot
Time-resolved graph of cycles for the polymorphic identification of the H-Bonded network in flexible (bio-)molecules
J. Chem. Theory Comput. 20:1019-1035 (2024)
DOI : 10.1021/acs.jctc.3c01031
5- A. Hashemi, S. Bougueroua, M.-P. Gaigeot, E.A. Pidko
HiREX : High-Throughput Reactivity Exploration for Extended Databases of Transition Metal Catalysts
J. Chem. Inf. Model. 63 : 6081-94 (2023)
DOI : 10.1021/acs.jcim.3c00660
6- A. Hashemi, S. Bougueroua, M.-P. Gaigeot, E.A. Pidko
ReNeGate : A Reaction Network Graph-Theoretical Tool for Automated Mechanistic Studies in Computational Homogeneous Catalysis
J. Chem. Theory Comput. 18 :7470-7482 (2022)
DOI : 10.1021/acs.jctc.2c00404
7- A. Serva, S. Pezzotti, S. Bougueroua, D. Galimberti, M.-P. Gaigeot
Combining ab-initio and classical molecular dynamics simulations to unravel the structure of the 2D-HB-Network at the air-water interface.
J. Mol. Struct., 1165:71-78 (2018)