A Suite of Foundation Models Captures the Contextual Interplay Between Codons.

Area of Science:

• Computational Biology
• Genomics
• Molecular Biology

Background:

• The genetic code assigns multiple codons to many amino acids (synonymous codons).
• Synonymous codon choice is non-random and influences protein function and cellular regulation.
• Current protein foundation models overlook the regulatory role of coding sequences.

Purpose of the Study:

• Introduce cdsFM, a suite of codon-resolution large language models (EnCodon and DeCodon), to address the gap in understanding synonymous codon function.
• Evaluate the models' ability to learn codon-amino acid relationships and predict synonymous codon usage.
• Assess the impact of synonymous codon changes on protein expression and model performance.

Main Methods:

• Developed and pre-trained cdsFM (up to 1B parameters) on 60 million protein-coding sequences from over 5,000 species.
• Generated a dataset measuring protein expression levels under varying synonymous codon contexts.
• Fine-tuned EnCodon models to predict consequences of synonymous codon choices and applied them to clinical datasets.

Main Results:

• cdsFM models effectively recapitulate the genetic code structure and outperform existing genomic models in zero-shot and few-shot learning tasks.
• Larger cdsFM models demonstrate superior prediction of synonymous codon choice.
• Fine-tuned EnCodon models accurately predict the impact of synonymous codon changes on protein expression and identified potentially pathogenic synonymous variants.

Conclusions:

• The cdsFM suite provides a powerful tool for decoding the functional grammar of synonymous codon usage.
• EnCodon models can be fine-tuned to predict the contextual effects of synonymous codon variations.
• The study identifies numerous likely pathogenic synonymous codons, with experimental validation, highlighting their clinical relevance.