A comparative study of parallel metaheuristics executed in grid environments is proposed, having as case study a genetic algorithm, a simulated annealing algorithm and a random search method. The random search method was constructed in order to offer a lower bound for the comparison. Furthermore, a conjugated gradient local search method is employed for each of the algorithms, at different points on the execution path. The algorithms are evaluated using the protein structure prediction problem, the benchmark instances consisting of the tryptophan-cage protein (Brookhaven protein data bank ID 1L2Y) and alpha-cyclodextrin. The algorithms are designed to benefit from the grid environment although having no particular optimization for the specified benchmarks. The presented results are obtained by running the algorithms independently and, in a second time, in conjunction with the conjugated gradient search method. Experimentations were performed on a nation-wide grid reuniting five distinct administrative domains and cumulating 400 CPUs. The complexity of the protein structure prediction problem remains prohibitive as far as large proteins are concerned, making the use of parallel computing on the computational grid essential for its efficient resolution.