RNA language models predict mutations that improve RNA function.

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• Structured RNA is crucial for biological processes like gene expression and catalysis.
• Current RNA structure prediction is limited by a lack of phenotype-associated reference data.
• Understanding RNA function requires linking structure to organismal traits.

Purpose of the Study:

• To develop GARNET, a novel database for RNA structural and functional analysis.
• To link RNA sequences with environmental temperatures from the Genome Taxonomy Database (GTDB).
• To create sequence- and structure-aware RNA generative models.

Main Methods:

• Developed GARNET by integrating RNA data with GTDB organismal temperatures.
• Implemented overlapping triplet tokenization for optimal encoding in GPT-like models.
• Utilized hyperthermophilic RNAs and generative models to identify beneficial mutations.

Main Results:

• GARNET provides a unique resource linking RNA to environmental growth temperatures.
• Developed effective RNA generative models leveraging sequence and structure information.
• Identified specific mutations in ribosomal RNA enhancing thermostability in E. coli.

Conclusions:

• GARNET and the developed deep learning models offer a foundation for RNA research.
• This work advances understanding of the RNA sequence-structure-function-phenotype relationship.
• The findings facilitate the design of more stable RNA molecules for biotechnological applications.