We have developed a new de novo protein structure prediction algorithm, SKIFOLDING. The search space of the algorithm is built from all the possible combinations of secondary structure elements in a given subset of protein structures databank. Search directions follow the best possible folding pathways, with a simple statistical score function. We tested the algorithm on the ability to distinguish between different families within the immunoglobulin-like type of fold, for 9 proteins with known structure. For the most of tested proteins the correct structure was presented in the program output and false-positive decoys were of worse quality than the correct structure.