Short human eccDNAs are predictable from sequences.

This article focuses on the predictability of short extrachromosomal circular DNAs (eccDNAs) in human cells using DNA sequence-based deep learning models. Despite previous views suggesting that short eccDNAs are randomly generated due to their widespread and diverse genomic origins, the study developed DeepCircle, a bioinformatics framework employing convolutional neural networks (CNN) and Bidirectional Encoder Representations from Transformers (BERT) adapted for DNA (DNABERT), to demonstrate that short eccDNAs share intrinsic sequence features enabling accurate prediction across multiple datasets from different tissue and cell line origins. The models achieved approximately 80% accuracy, revealing common sequence characteristics such as higher GC content and specific dinucleotide frequencies, and identified motifs related to zinc-finger protein binding, suggesting potential biological roles. These findings indicate that while eccDNAs originate from diverse genomic locations, their formation propensity is encoded in DNA sequences, providing a foundation for future research into eccDNA biogenesis and biomarker development.