ADAR-GPT: A continually fine-tuned language model for predicting A-to-I RNA editing sites.

Area of Science:

• Molecular Biology
• Genomics
• Bioinformatics

Background:

• Adenosine-to-inosine (A-to-I) RNA editing, catalyzed by ADAR enzymes, is crucial for transcript regulation and therapeutic development.
• Predicting specific RNA editing sites remains a significant challenge in the field.

Purpose of the Study:

• To introduce ADAR-GPT, a flexible framework for predicting A-to-I RNA editing sites using GPT-class language models.
• To evaluate ADAR-GPT's performance against existing computational methods.

Main Methods:

• ADAR-GPT employs a model-agnostic fine-tuning approach on sequence context around candidate adenosine sites.
• A two-stage continual fine-tuning strategy was used with GTEx liver data, incorporating lower editing thresholds as curriculum data.
• The model was benchmarked against convolutional and foundation model architectures.

Main Results:

• ADAR-GPT achieved competitive or superior performance compared to established computational approaches.
• The model demonstrated an improved balance of recall, precision, and specificity, with strong operating-curve metrics.
• The framework is reproducible and portable across different GPT backbones.

Conclusions:

• ADAR-GPT offers a robust and adaptable method for classifying RNA editing sites.
• The tool provides practical adenosine scoring to prioritize experimental targets and guide RNA design.
• This framework can be readily adapted for new datasets and advanced model architectures.