In this paper, we adopt a position specific scoring matrix as an abstraction of amino acid type to derive two new statistical potentials for protein structure prediction, and investigated its effect on the quality of the potentials compared to that derived using residue specific amino acid identity. For stringent test of the potential quality, we carried out folding simulations of 91 residue A chain of protein 2gpi, and found unexpectedly that the abstract amino acid type improved the quality of the one-body type statistical potential, but not for the two-body type statistical potential which describes long range interactions. This observation could be effectively used when one develops more accurate potentials for structure prediction, which are usually involved in merging various one-body and many-body potentials.