codonGPT: reinforcement learning on a generative language model enables scalable mRNA design.

Area of Science:

• Computational Biology
• Bioinformatics
• Synthetic Biology

Background:

• Reinforcement learning (RL) is established in engineering design but underexplored in biological applications.
• Current generative language models for biology primarily focus on DNA, RNA, or proteins, neglecting messenger RNA (mRNA).
• The lack of generative models for mRNA impedes scalable design for therapeutics and synthetic biology.

Purpose of the Study:

• To introduce the first generative language model, codonGPT, specifically trained on mRNA sequences.
• To develop a method for mRNA design as a constrained language modeling task.
• To demonstrate mRNA sequence optimization using RL and codonGPT.

Main Methods:

• Trained codonGPT exclusively on 338,417 mRNA sequences from model organisms.
• Introduced an inference-time masking method to constrain synonymous sequences for mRNA.
• Applied reinforcement learning (RL) to codonGPT for optimizing mRNA sequences based on biological constraints (expression, stability, GC-content).

Main Results:

• Developed codonGPT, the first generative language model for mRNA.
• Demonstrated successful mRNA sequence optimization for HLA-A and ACTB genes.
• Validated the approach for optimizing reporter genes like GFP and beta-lactamase.

Conclusions:

• codonGPT represents a significant advancement in generative modeling for mRNA.
• The RL-based optimization framework enables tailored mRNA design for various biological applications.
• This work paves the way for enhanced therapeutics, synthetic biology, and protein engineering through optimized mRNA sequences.