CodonTransformer: a multispecies codon optimizer using context-aware neural networks.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Genetic code degeneracy allows multiple DNA sequences for one protein.
• Optimizing codon usage is computationally complex due to combinatorial explosion.
• Natural sequences offer insights into evolutionary codon usage rules.

Purpose of the Study:

• To develop a deep learning model for optimizing DNA sequences across multiple species.
• To generate host-specific DNA sequences with natural codon distribution and minimal regulatory elements.

Main Methods:

• Developed CodonTransformer, a multispecies deep learning model using a Transformer architecture.
• Trained the model on over 1 million DNA-protein pairs from 164 organisms.
• Implemented a sequence representation strategy combining organism, amino acid, and codon encodings.
• Introduced the Shared Token Representation and Encoding with Aligned Multi-masking (STREAM) strategy.

Main Results:

• CodonTransformer demonstrates context-aware sequence generation.
• Generated DNA sequences exhibit natural-like codon distribution profiles.
• Sequences generated have minimal negative cis-regulatory elements.
• The model is adaptable for host-specific codon optimization.

Conclusions:

• CodonTransformer provides an effective framework for codon optimization.
• The open-access model and Colab interface facilitate broader application.
• This approach leverages machine learning to address challenges in synthetic biology and genetic engineering.