PlasmidGPT: A generative framework for plasmid analysis and generation.

Area of Science:

• Synthetic Biology
• Bioinformatics
• Computational Biology

Background:

• Engineered plasmids are crucial tools in molecular biology and biotechnology.
• Analyzing and designing plasmid sequences is complex and computationally intensive.
• Existing models may not fully capture the nuances of plasmid sequence information.

Purpose of the Study:

• Introduce PlasmidGPT, a generative language model for plasmid sequence analysis and design.
• Leverage informative sequence embeddings for visualization and diversity analysis.
• Achieve state-of-the-art performance in predicting plasmid features and origins.

Main Methods:

• Pretraining PlasmidGPT on a large dataset of 153,208 engineered plasmid sequences from Addgene.
• Utilizing learned sequence embeddings for visualization and diversity analysis.
• Evaluating PlasmidGPT's performance in feature prediction, lab-of-origin prediction, and host taxonomy prediction.

Main Results:

• PlasmidGPT learns informative embeddings for visualizing research topics and analyzing plasmid diversity.
• Achieved state-of-the-art performance in lab-of-origin prediction for engineered plasmids.
• Demonstrated generalization to natural plasmids for host taxonomy prediction (phylum and genus level).
• Enabled controlled generation of functional plasmid sequences with accurate part co-occurrence and synteny.

Conclusions:

• PlasmidGPT offers a powerful tool for understanding and designing plasmid sequences.
• The model's embeddings facilitate novel insights into plasmid research and diversity.
• PlasmidGPT advances capabilities in synthetic biology and microbial genomics through accurate prediction and controlled generation.