Transformer-based operon prediction using textual representations of gene pairs.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Operons are key bacterial gene regulatory units, crucial for understanding genome organization and gene co-expression.
• Computational operon prediction aids pathway reconstruction and comparative genomics but faces limitations with existing tools.

Purpose of the Study:

• To develop a scalable and broadly applicable computational method for operon prediction.
• To leverage natural language processing techniques for improved operon identification in bacterial genomes.

Main Methods:

• A transformer-based approach reformulating operon prediction as a binary text classification task for adjacent gene pairs.
• Serialization of genomic features (orientation, distance, GC content, annotations, protein families, conservation) into natural language descriptions.
• Utilizing a RoBERTa-based model for operon classification.

Main Results:

• The RoBERTa model achieved competitive predictive performance across multiple evaluation settings, including leave-one-species-out analysis.
• Demonstrated that sequence-derived and annotation-based features are sufficient for high performance.
• Showed stable performance even when selected features were removed during inference.

Conclusions:

• Transformer-based models offer a flexible and effective approach for bacterial operon prediction.
• The proposed method overcomes limitations of existing tools, particularly for poorly annotated genomes.
• This approach facilitates downstream genomic analyses by improving operon identification accuracy.