RNA language models predict mutations that improve RNA function.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Biology

Background:

• RNA structure is crucial for biological processes but difficult to predict due to limited data.
• Existing RNA sequence data often lacks organismal phenotypes, hindering functional analysis.

Purpose of the Study:

• To create a comprehensive RNA database (GARNET) linking genomic data with environmental temperatures.
• To develop advanced RNA generative models for structure and function prediction.
• To identify RNA mutations that enhance thermostability.

Main Methods:

• Developed GARNET database by integrating RNA sequences from GTDB genomes with organismal growth temperatures.
• Constructed deep and diverse RNA sequence alignments for machine learning.
• Developed a GPT-like language model for RNA using overlapping triplet tokenization.
• Leveraged hyperthermophilic RNAs and generative models to identify stabilizing mutations.

Main Results:

• Established GARNET, a novel resource for RNA structural and functional analysis.
• Defined requirements for sequence- and structure-aware RNA generative models.
• Created an effective RNA language model with optimized tokenization.
• Identified specific mutations in ribosomal RNA conferring increased thermostability.

Conclusions:

• GARNET provides a foundation for understanding RNA sequence-structure-function relationships.
• The developed deep learning models advance RNA structure prediction and design.
• This work facilitates the study of RNA in diverse environmental conditions and applications.