Design and classify innovative antimicrobial and dual- and multi-functional peptides using generative artificial intelligence

Antimicrobial resistance is a growing global crisis, where antimicrobial peptides (AMPs) have emerged as promising alternatives to conventional antibiotics due to their unique mechanisms of action. However, traditional peptide design approaches, especially when designing multi-functional peptides, are often costly and time-consuming. This study explores the potential of ChatGPT as a prompt-based tool used without additional fine-tuning, for the classification and design of peptides with varying degrees of functional capabilities. We optimized the AMP design prompts by incorporating relevant information and refining them to design multi-functional peptides. For peptide classification, representative examples of positive and negative sequences were sourced from databases specific to each functionality, and a small exploratory set (10 positive sequences and 10 negative sequences) from the respective databases were randomly selected. In this preliminary evaluation, ChatGPT designed single- and dual-functional peptides with a promising apparent performance (accuracy > 0.9) and a decreased performance for triple- and quadruple-functional peptides (accuracy of 0.8) based on in silico testing. Additionally, ChatGPT showed promising preliminary results in classifying peptides with antimicrobial, non-hemolytic, cell-penetrating, and anticancer properties. Incorporating in-context example sequences significantly improved its classification performance. This work provides a proof-of-concept study of using a generative artificial intelligence tool for classification and multi-functional peptide design. While experimental validation is required, these findings suggest that accessible, non-coding Large Language Model approaches may support early-stage in silico peptide discovery that may be used in combination with existing computational tools.