Enhancing Antimicrobial Peptide Functionality and Manufacturability Through Deep Learning-Based Sequence Design

Protein sequence design aims to create proteins with enhanced functionality, manufacturability compared to their natural counterparts. While deep learning methods have shown success, they often neglect critical functional aspects of proteins. This study introduces an intuitive optimization approach to improve properties like solubility, flexibility, stability, while preserving protein structure. This method represents a significant advancement in developing synthetic antimicrobial peptides. The initial peptides designed include "Defensin, alpha 5," found in the human intestine and known for its antimicrobial properties. Additionally, "Nisin A," a bacterial peptide effective against Gram-positive bacteria widely used as a food preservative, and "Plectasin," a fungal peptide noted for its activity against Gram-positive bacteria, were designed using this approach. We propose a deep learning-based method for optimizing complete protein design. Initially, a protein sequence undergoes substituting aa residues by a sequence design model (i.e. ProteinMPNN). The designed sequence's solubility, flexibility, and instability indices are then computed. If the sequence surpasses the current solution or meets the Metropolis criterion, it is accepted. Accepted sequences undergo further optimization by redesigning other randomly chosen positions, iteratively, for a set number of cycles. Peptides designed using our approach exhibit enhanced biophysical properties compared to natural peptides (Table 1). Ensuring the preservation of their 3D structures is crucial. We utilized OmegaFold to predict and compare structures of natural and designed sequences (Figure 1). The findings indicate that peptides designed with this method closely resemble natural peptide sequences and structures. This approach facilitates the creation of highly potent antimicrobial peptides in laboratory settings.