This study presents a novel approach combining protein structure prediction with Bayesian Optimization (BO) for peptide design. The goal is to identify peptides with unique biological properties by leveraging deep learning and Bayesian Optimization. The method uses objective functions to guide and enhance the optimization of new peptide sequences, aiming to generate peptides with optimal biological properties. The research integrates ColabFold for protein structure prediction with BO to refine peptide sequences. By using latent embeddings, the method efficiently explores the sequence space, prioritizing promising candidates through model estimates and evaluations. The study also incorporates specific scoring functions to evaluate complex structures, enhancing the optimization of peptide sequences. The methodology was validated using native protein-peptide complexes from PDB, with artificial mutations to test the model's ability to recover native structures. The results show that the method effectively identifies peptides close to their native counterparts, with improved stability and binding affinity. The study highlights the potential of combining BO with protein modeling for peptide design, although it acknowledges computational limitations and the need for further improvements in embedding techniques.This study presents a novel approach combining protein structure prediction with Bayesian Optimization (BO) for peptide design. The goal is to identify peptides with unique biological properties by leveraging deep learning and Bayesian Optimization. The method uses objective functions to guide and enhance the optimization of new peptide sequences, aiming to generate peptides with optimal biological properties. The research integrates ColabFold for protein structure prediction with BO to refine peptide sequences. By using latent embeddings, the method efficiently explores the sequence space, prioritizing promising candidates through model estimates and evaluations. The study also incorporates specific scoring functions to evaluate complex structures, enhancing the optimization of peptide sequences. The methodology was validated using native protein-peptide complexes from PDB, with artificial mutations to test the model's ability to recover native structures. The results show that the method effectively identifies peptides close to their native counterparts, with improved stability and binding affinity. The study highlights the potential of combining BO with protein modeling for peptide design, although it acknowledges computational limitations and the need for further improvements in embedding techniques.