Advancements in Protein Engineering through AI: Raygun System

Researchers at Duke University, led by Rohit Singh, have developed an artificial intelligence (AI) system called Raygun that can miniaturize, enlarge, and extensively rewrite proteins, as detailed in their study published in the journal Nature. This innovation holds significant implications for gene therapy and protein research.

Significance of Protein Miniaturization

• Gene Therapy:
  • Gene therapy involves delivering a working copy of a gene into a patient's cells, often using viral shells with limited capacity.
  • Miniaturizing proteins allows gene therapies to fit into these constrained spaces, potentially reducing costs and increasing accessibility.
  • Examples of expensive gene therapies include Hemgenix ($3.5 million) for haemophilia and Zolgensma ($2.1 million) for spinal muscular atrophy.
• Research Applications:
  • Smaller proteins can be more efficiently attached to reporters like the green fluorescent protein for tracking without distortion.

Understanding Proteins as Language

• Proteins are like sentences constructed from 20 amino acids, determining their shape and function.
• Genes, encoding proteins, use a four-nucleotide alphabet, drawing a parallel to language.
• Protein language models learn patterns from millions of sequences, establishing a grammar that links to biological structure and function.

Functionality of Raygun

• Raygun leverages protein language models to condense proteins into fixed summaries, allowing for modifications without altering functionality.
• It utilizes a probability distribution to generate protein variants swiftly, outperforming conventional diffusion-based methods.
• Users can adjust the degree of deviation from the original protein and the desired length of the result.

Experimental Outcomes

• Raygun was tested on a range of proteins, successfully shortening mTOR by over 500 amino acids while maintaining structure.
• Generated variants of fluorescent proteins like eGFP and mCherry were tested, with six out of eight glowing when expressed in human cells.
• The smallest functional variants were shorter than 96% of known fluorescent proteins, demonstrating the system's efficacy.
• Proteins tolerated significant edits, surpassing traditional editing methods that fail with minimal alterations.

Current Limitations and Future Directions

• Current variants, while functional, are dim compared to established laboratory standards, necessitating further evolution.
• The AI's reliance on evolutionary data limits its ability to preserve engineered properties added by scientists.
• Ethical considerations are paramount, highlighted by adherence to the Responsible AI for Biodesign principles.