Molecular dynamics (MD) simulation has become a powerful tool for studying the structures and functional mechanisms of biomolecules, and its reliability crucially depends on the accuracy of underlying force fields. This perspective describes our recent efforts to develop more accurate protein force fields by improving the description of intrinsic conformational preferences of amino acid residues using residue-specific dihedral-angle-related parameters. Both backbone and side-chain conformational distributions and their coupling were optimized to fit those from a protein coil library. The resulting force fields RSFF1 and RSFF2 have been found to be more accurate than popular protein force fields, in reproducing experimental structural data of various peptides and proteins. They have also been successfully used in studying folding mechanisms and refinement of structure models. Further methodology developments related to intrinsically disordered proteins (RSFF2+) and a more universal implementation (RSFF2C) based on CMAP potentials are also described.