D&D1 - Organic Crystals

Molecular materials, e.g. those formed by pharmaceutical compounds, pose formidable modeling challenges.  This D&D project has developed, demonstrated, and distributed breakthrough simulation techniques, powered by accurate quantum calculations, advanced statistical sampling and machine learning, to address these challenges, increasing the predictive power of simulations and bringing together computation and experiments.


The project D&D1 — Understanding Complex Molecular Crystals: Structures and Properties — unites the expertise of several groups within the MARVEL collaboration, and leverages some of the outcomes of the development work of phase I, to bring about a new level of understanding of molecular crystals. By combining physics-based and data-driven modeling this project is demonstrating how larger and more flexible compounds, such as those that make up a large fraction of last-generation drugs, can be modeled with high accuracy, describing their stability and their properties in a way that complements and augments experimental investigations.


The project is led by Michele Ceriotti.

Group Leaders

Michele Ceriotti
Project leader
EPFL, Lausanne
Clémence Corminboeuf
Group leader
EPFL, Lausanne
Lyndon Emsley
Group leader
EPFL, Lausanne
Stefan Goedecker
Group leader
UniBas, Basel
Michele Parrinello
Group leader
USI, Lugano
Sereina Riniker
Group leader
ETHZ, Zurich

Median-error binding curves for six different classes of intermolecular interactions. (Black lines) quantum-mechanical calculations. (Green lines) predictions of a conventional ML model, that relies entirely on short-range interactions. (Blue lines) multi-scale LODE model, that combine short-range information with an asymptotic term mimicking multipole electrostatics [1].

Machine-learned prediction of the charge density of the peptide encephalin [2]. The charge is sufficiently accurate to capture the fine details of the charge density that are associated with covalent and non-covalent interactions, as identified by the DORI fingerprint. The error (shown in the bottom panels for enkephalin and a cyclopeptide) concentrates in the region of the peptide bond, as the training set contains no examples of such chemical feature.


[1] Grisafi-arxiv-2020 A. Grisafi, J. Nigam, and M. Ceriotti, Multi-scale approach for the prediction of atomic scale properties, arxiv:2008.12122 (2020). [Open Access URL]
 

[2] A. Fabrizio, A. Grisafi, B. Meyer, M. Ceriotti, C. Corminboeuf, Electron density learning of non-covalent systemsChemical Science 10, 9424 (2019). [Open Access URL] 


Related publications (until January 2020)