Protein structure prediction is a challenging optimisation problem to the computer scientists. A large number of ex- isting (meta-)heuristic search algorithms attempt to solve the problem by exploring possible structures and finding the one with minimum free energy. However, these algo- rithms often get stuck in local minima and thus perform poorly on large sized proteins. In this paper, we present a random-walk based stagnation recovery approach. We tested our approach on tabu-based local search as well as population based genetic algorithms. The experimental results show that, random-walk is very effective for escaping from local minima for protein structure prediction on face- centred-cubic lattice and hydrophobic-polar energy model.