Systematic construction of coarse-grained molecular models from detailed atomistic simulations, and even from ab initio simulations is discussed. Atomistic simulations are first performed to extract structural information about the system, which is then used to determine effective potentials for a coarse-grained model of the same system. The statistical-mechanical equations expressing the canonical properties in terms of potential parameters can be inverted and solved numerically according to the iterative Newton scheme. In our previous applications, known as the Inverse Monte Carlo, radial distribution functions were inverted to reconstruct pair potential, while in a more general approach the targets can be other canonical averages. We have considered several examples of coarse-graining; for the united atom water model we suggest an easy way to overcome the known problem of high pressure. Further, we have developed coarse-grained models for L- and D-prolines, dissolved here in an organic solvent (dimethylsulfoxide), keeping their enantiomeric properties from the corresponding all-atom proline model. Finally, we have revisited the previously developed coarse-grained lipid model based on an updated all-atomic force field. We use this model in large-scale meso-scale simulations demonstrating spontaneous formation of different structures, such as vesicles, micelles, and multi-lamellar structures, depending on thermodynamical conditions.