Peptide-GPT: Generative Design of Peptides using Generative Pre-trained Transformers and Bio-informatic Supervision • PeptideGPT introduces a protein language model designed for generating peptide sequences with specific properties: hemolytic, non-hemolytic, non-fouling, and soluble. Unlike traditional protein design methods, PeptideGPT leverages fine-tuned generative pre-trained transformers to achieve tailored de novo protein generation. • The model was fine-tuned from ProtGPT2 using specialized datasets for each targeted property. Generated sequences go through a rigorous evaluation pipeline, including perplexity ranking, bioinformatic supervision for protein validity, and structure prediction using ESMFold to ensure stable and ordered structures. • PeptideGPT achieved accuracy rates of 76.26% for hemolytic, 72.46% for non-hemolytic, 78.84% for non-fouling, and 68.06% for soluble protein generation. These accuracies indicate that PeptideGPT can effectively learn the relationship between amino acid sequences and their properties, providing reliable protein design. • A key innovation is the integration of classifiers like PeptideBERT and HAPPENN in the evaluation pipeline to predict whether the generated sequences possess the desired properties. This classifier-based validation helps confirm that the output proteins are biologically relevant and useful. • The authors also utilized a convex hull approach to filter valid proteins, ensuring that generated sequences are not only syntactically correct but also biologically feasible. • Compared to other models like ProtGPT2, PeptideGPT shows a significant improvement in generating proteins with the desired properties, demonstrating the effectiveness of the property-specific fine-tuning process. • The multi-step validation, including perplexity and structure checks, highlights PeptideGPT’s robustness, making it a promising tool for synthetic biology, drug development, and therapeutic applications. 💻Code: github.com/aayush-shah14/Pep… 📜Paper: arxiv.org/abs/2410.19222 #peptidedesign #proteinengineering #machinelearning #syntheticbiology #GPT #bioinformatics #drugdiscovery #transformers #PeptideGPT