Protein design is a core challenge in biotechnology, aiming to design novel protein sequences with specific functions. While advances have been made in deep generative models that design function-based proteins based on textual descriptions, ensuring structural validity has been challenging. This study explored ways to improve the foldability of generative models by incorporating natural protein fragments, drawing on existing protein design methods that utilize natural protein structures. We demonstrated that random fragment incorporation alone improved foldability, and based on this, we proposed a novel protein design approach, ProDVa. ProDVa integrates a text encoder for functional descriptions, a protein language model for protein design, and a fragment encoder that dynamically retrieves protein fragments based on textual descriptions. Experimental results demonstrate that ProDVa effectively designs functionally aligned and structurally valid protein sequences, achieving a similar level of functional alignment with a smaller training data set compared to state-of-the-art models while designing more well-folded proteins.
