Circular RNA identification using a genomic language model and a small number of authenticated examples

Genomic language models (gLMs) hold great promise for deciphering biological sequences, yet their effectiveness is hindered by the limited number of experimentally verified examples available for model training, a ubiquitous bottleneck for supervised machine learning. To overcome this challenge, we developed circFormer, the first gLM-driven approach for circular RNA (circRNA) identification. circFormer integrates curriculum learning with gLM fine-tuning: a Nucleotide Transformer model is first trained on a small set of validated circRNAs, the resulting model is used as a teacher to score [~]2.3 million noisy candidates, and the model is then fine-tuned with the noisy candidates along with their scores to improve prediction. Operating either as a standalone predictor or as a filter for existing pipelines, circFormer consistently outperformed traditional machine-learning approaches in accuracy and robustness. Among 50 circFormer-selected candidates that were overlooked by most existing tools, experimental validation using RNase R digestion and RT-qPCR confirmed 94.1% (32/34) of the evaluable candidates as genuine circRNAs. To enhance interpretability, we introduced a model-agnostic, dual-level explainable AI strategy that reveals mechanistic signatures of circRNA formation. circFormer provides a scalable, interpretable, and generalizable framework for converting noisy high-throughput data into reliable functional annotations, highlighting a practical path forward for gLM-based genomics in data-scarce settings.