Project

molecular dynamics; Protein Engineering; artificial intelligence; multiscale modelling; Photovoltaics; machine learning; Biomimetic Design; Computational Chemistry; QM/MM ; Computational Biophysics

Lead
Die nächste Generation von first-principles Molecular Dynamik Simulationen mit Anwendungen in der Biomimetik und im Design von Materialien
Lay summary
Während unseres bisherigen SNF-Grants haben wir Multi-Rang-QML/QMH MD-Methoden entwickelt, die es erlauben, Simulationen unter Beibehalt der rechnerischen Kosten einer niederrangigen Methode mit der Genauigkeit einer hochrangigen Methode durchzuführen. Hierzu wurden komplett zeitreversible Multi-Zeitschritt (MTS) Integratoren verwendet. Durch die Verbindung des Programms CPMD mit anderen quantenchemischen Programmen (GAUSSIAN und TURBOMOLE) waren wir in der Lage, bspw. B3LYP oder MP2-Rechnungen mit einem um einen Faktor 5-10 tieferen Rechenaufwand durchzuführen. Wir schlagen hier vor, die Flexibilität und Effizienz dieser MTS-QML/QMH Simulationen zu vergrössern (durch eine Erweiterung auf gemischte QM/MM- und QML&QMH/MM-Systeme; durch Implementierung stochastischer Thermostaten um Resonanzen zu minimieren; sowie die Entwicklung und Implementierung neuartiger Ansätze zur schnellen Berechnung von Austauschintegralen), und diese mit Ansätzen aus künstlicher Intelligenz (AI) zu kombinieren. Die Kombination von Kraft-adaptierten klassischen Kraftfeldern mit spontanem Maschinellen Lernen (ML) für die Evaluierung quantenmechanischer Kräfte sollte es ermöglichen, Simulationen der nächsten Generation mit um Grössenordnungen längeren Zeitschritten durchzuführen. Verschiedene Kombinationen dieser MTS-ML-MD-Ansätze (MTS-MM&QM(ML), MTS-QML(ML)-QMH, MTS-MM&QM(ML)/MM etc.) werden sich in einem breiten Anwendungsbereich in der Biologie und den Materialwissenschaften als äusserst nützlich erweisen. In diesem Antrag liegt der Schwerpunkt im Speziellen auf Anwendungen mit Bezug zu Peptid- und Proteindesign, sowie auf der Entwicklung biomimetischer Systeme für Katalyse und Lichtnutzung.

Direct link to Lay Summary

Last update: 03.05.2016

Active participation

Self-organised

Title	Year

Number	Title	Start	Funding scheme

Molecular dynamics (MD) simulations are arguably the most widely used molecular simulation method today and force-field based classical MD has been highly successful in the simulation of condensed phase systems for both biological and materials science applications. In addition, the advent of quantum mechanical based first-principles MD (FPMD) has enabled the treatment of electronic phenomena such as chemical reactions and photoexcitations. The introduction of multiscale mixed quantum mechanical/molecular mechanical (QM/MM) simulations in combination with a range of powerful enhanced sampling methods has allowed extending both spatial and temporal scales of FPMD. Together with the increase in computer power, it has become possible to perform FPMD with several 100 -1000 of atoms for 10-100 picoseconds. However, the combined challenge of system size, sampling time and high accuracy is still formidable. Ideally, one would like to perform MD simulations with the system size and sampling times typical of force-field based MD but with the accuracy of a high level ab initio method. During our previous SNF grant, we have developed multi-rung QML&QMH MD methods that allow to perform simulations with the accuracy of a high level method for essentially the cost of a lower level description. To this end, we have made extensive use of fully time-reversible multiple time step (MTS) integrators. By interfacing the FPMD program CPMD with other quantum chemical programs (GAUSSIAN and TURBOMOLE), we were able to perform e.g. B3LYP or MP2 based FPMD with a 5-10 times reduced computational cost. Here we propose, to further extend the versatility and efficiency of these MTS-QML&QMH simulations (via extension into a mixed QM/MM and QML&QMH/MM context; by implementation of stochastic thermostats to minimize resonances; and through the development and implementation of a novel approach for fast exact exchange) and to combine them with approaches from artificial intelligence. Combining force-matched classical force fields with on-the-fly machine learning (ML) for QM force evaluations, it should become feasible to run next-generation FPMD simulations with several orders of magnitude speedup.These MTS-ML MD approaches in various combinations (MTS-MM&QM(ML), MTS-QML(ML)-QMH, MTS-MM&QM(ML)/MM etc..) will be highly useful for a wide range of applications in biology and material science. In this proposal, we focus especially on applications related to peptide and protein engineering and the development of biomimetic systems for catalysis and light harvesting. The design of such biomimetic systems involves the steps of 1) FPMD or QM/MM simulations of the mechanism of action of the natural target; 2) analysis of the simulation data and identification of the relevant descriptor that influence the function of the natural system and 3) search of chemical and/or sequence space for optimal biomimetic counter parts. For all three steps, we will combine molecular simulations with approaches from artificial intelligence: For step 1), the use of an on-the-fly machine learning approach will enable highly efficient force evaluations that can be used as lower level forces within an MTS scheme. Feature selection and causality inference techniques will be crucial for an identification of the relevant descriptor in the highly dimensional simulation data generated in step 1) and last but not least, in step 3) efficient explorations of chemical and sequence space will be performed with the help of evolutionary optimization algorithms (genetic algorithms and particle swarm optimization) by using and further extending the program toolbox EVOLVE that we have been developing during our previous SNF grants.