Protein structure prediction is one of the most challenging problems in computational biology and remains unsolved for many decades. In a simplified version of the problem, the task is to find a self-avoiding walk with the minimum free energy assuming a discrete lattice and a given energy matrix. Genetic algorithms currently produce the state-of-the-art results for simplified protein structure prediction. However, performance of the genetic algorithms largely depends on the encodings they use in representing protein structures and the twin removal technique they use in eliminating duplicate solutions from the current population. In this paper, we present a new efficient encoding for protein structures. Our encoding is nonisomorphic in nature and results into efficient twin removal. This helps the search algorithm diversify and explore a larger area of the search space. In addition to this, we also propose an approximate matching scheme for removing near-similar solutions from the population. Our encoding algorithm is generic and applicable to any lattice type. On the standard benchmark proteins, our techniques significantly improve the state-of-the-art genetic algorithm for hydrophobic-polar (HP) energy model on face-centered-cubic (FCC) lattice.