Systematic coarse-graining using structural information : applications to lipid membranes

Within last 20 years, advances in computational power and methodology have made computer simulations an integral part of studies of biomolecular systems. Simulations on all-atom level are routinely used to study, e.g., microscopic details of lipid aggregates and proteins. However, many phenomena are still outside the reach of all-atom simulations, and coarser models are needed. Detailed information from all-atom models can provide input data for parameterizing coarse-grained (CG) models. Techniques for such parameterization are called systematic coarse-graining methods, and can be based, e.g., on matching forces or structural information between the two resolutions. The main part of this dissertation employs inverse Monte Carlo (IMC) for constructing CG models for a lipid membrane containing dipalmitoylphosphatidylcholine (DPPC) and cholesterol. Three 2D models are constructed at different levels of resolution, in each case matching the radial distribution functions (RDFs) of the CG model to those from atom-scale simulations. The main results are the presence of cholesterol-rich and cholesterol-poor domains at intermediate cholesterol concentrations and the presence of strong tail density fluctuations at low cholesterol concentrations. The former agrees with the experimental studies of the system, while the latter was confirmed through atom-scale simulations. Accurate quantitative studies were restricted by transferability problems in all the CG models; hence, focus is on comparing the different models and critical discussion of the RDF inversion as a basis for coarse-graining. The IMC method is also improved by increasing its tolerance to statistical noise, as well as through inclusion of a virial pressure constraint and generalization to models where particles have internal degrees of freedom. The dissertation also discusses the analysis of individual lipid conformations from atom-scale simulations using self-organizing maps (SOMs), as well as the use of SOMs in coarse-graining. Atomistic simulations provide a vast amount of data, and direct analysis may be difficult. SOM, an unsupervised machine learning method, is studied as an alternative to more traditional analysis. Focus is on determining good parameters for the method and on qualitative analysis based on the good visualization properties of SOM. The internal dynamics of the molecules are also analyzed using SOMs for visualization. A bilayer of palmitoyllinoleoyl-PC (PLPC) is used as a model system