Abstract The Protein Structure Prediction (PSP) problem is one of the most significant open problems in bioinformatics. In the AB off-lattice model, the protein sequence is labeled as ‘A’ or ‘B’ according to the amino acid classification of being hydrophobic or hydrophilic. It has been widely explored in the literature because polarity is one of the main driving forces behind protein structure definition. This work provides a high-performance hybrid algorithm to approach the 3D-AB off-lattice model through Graphics Processing Units (GPUs). The proposed hybrid algorithm, named cuHjDE–3D, is a self-adaptive Differential Evolution (DE) that uses the jDE mechanism to self-adapt the DE parameters and employs the Hooke-Jeeves Direct Search (HJDS) as the exploitation routine. The experiments were conducted on real protein sequences from the Protein Data Bank (PDB) and compared against state-of-the-art algorithms from the related literature concerning the 3D-AB off-lattice model. Moreover, we provide a methodology to compare a 3D-AB predicted conformation with its native conformation from the PDB repository using the RMSD metric. The obtained results highlight the optimization potential of the proposed method. Also, the GPU running time analysis reports the positive impact of using a massively parallel architecture, with speedups up to 277× , promoting the necessary scalability to handle the 3D-AB model.