Automatic Optimization of Lipid Models in the Martini Force Field Using SwarmCG

After two decadesof continued development of the Martinicoarse-grainedforce field (CG FF), further refinment of the already rather accurateMartini lipid models has become a demanding task that could benefitfrom integrative data-driven methods. Automatic approaches are increasinglyused in the development of accurate molecular models, but they typicallymake use of specifically designed interaction potentials that transferpoorly to molecular systems or conditions different than those usedfor model calibration. As a proof of concept, here, we employ SwarmCG, an automatic multiobjective optimization approachfacilitating the development of lipid force fields, to refine specificallythe bonded interaction parameters in building blocks of lipid modelswithin the framework of the general Martini CG FF. As targets of theoptimization procedure, we employ both experimental observables (top-downreferences: area per lipid and bilayer thickness) and all-atom moleculardynamics simulations (bottom-up reference), which respectively informon the supra-molecular structure of the lipid bilayer systems andon their submolecular dynamics. In our training sets, we simulateat different temperatures in the liquid and gel phases up to 11 homogeneouslamellar bilayers composed of phosphatidylcholine lipids spanningvarious tail lengths and degrees of (un)saturation. We explore differentCG representations of the molecules and evaluate improvements a posteriori using additional simulation temperatures anda portion of the phase diagram of a DOPC/DPPC mixture. Successfullyoptimizing up to & SIM;80 model parameters within still limitedcomputational budgets, we show that this protocol allows the obtainmentof improved transferable Martini lipid models. In particular, theresults of this study demonstrate how a fine-tuning of the representationand parameters of the models may improve their accuracy and how automaticapproaches, such as SwarmCG, may be very useful tothis end.

Automated summary

By using the SwarmCG automatic multiobjective optimization approach, we have successfully refined the bonded interaction parameters in the Martini coarse-grained force field (CG FF) lipid models. Through experimental observables and all-atom molecular dynamics simulations, we optimized 80 model parameters and obtained improved and transferable Martini lipid models.