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

Motivation: Operons are fundamental units of gene regulation in bacteria and can provide valuable insights into genome organization, co-expression, and functional relationships between genes. Computational prediction of operons can support downstream analyses such as pathway reconstruction, comparative genomics, and gene function inference. However, many existing tools rely on rigid feature engineering or curated interaction networks, limiting scalability and applicability to poorly annotated genomes. Results: We propose a transformer-based approach that reformulates operon prediction as a binary text classification task over adjacent gene pairs. By serializing genomic features, including gene orientation, intergenic distance, GC content, functional annotations, protein families, and conservation, into natural language descriptions, we enable pre-trained language models to perform operon classification using flexible, widely available inputs. A RoBERTa-based model achieves competitive predictive performance under multiple evaluation settings, including leave-one-species-out analysis over six bacterial genomes and benchmark comparisons on standard datasets. Through ablation and inference-time resilience analyses, we demonstrate that sequence-derived and annotation-based features are sufficient for competitive performance, and that performance remains stable when selected features are removed at test time. Availability and implementation: All datasets required to reproduce this study are publicly available at GitHub repository. Genome annotations were retrieved through the BV-BRC platform.