Short human eccDNAs are predictable from sequences.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Short extrachromosomal circular DNAs (eccDNAs) are abundant in eukaryotic cells, but their formation mechanisms remain debated.
• Previous studies suggested random or near-random eccDNA origins, hindering biomarker development.
• Recent interest in eccDNAs has spurred research into their specific formation patterns.

Purpose of the Study:

• To investigate the predictability of short eccDNA formation using deep learning.
• To challenge the notion of random eccDNA origins by identifying underlying sequence features.
• To develop a computational framework for analyzing eccDNA predictability.

Main Methods:

• Developed DeepCircle, a bioinformatics framework utilizing convolution- and attention-based neural networks.
• Applied DeepCircle to analyze human eccDNA datasets.
• Trained models to predict eccDNA formation based on DNA sequence features.

Main Results:

• DeepCircle achieved high prediction accuracy for short human eccDNAs across datasets (convolutional models: 79.65±4.7%, attention-based models: 83.31±4.18%).
• Identified shared DNA sequence features predictive of eccDNA formation, despite low similarity in genomic locations.
• Demonstrated that eccDNA predictability is encoded within their sequences, irrespective of tissue origin.

Conclusions:

• The formation of short eccDNAs is intrinsically predictable and encoded in their DNA sequences.
• Deep learning models can uncover hidden similarities in genomic data, re-evaluating perceived lack of specificity.
• Findings support further investigation of eccDNAs as potential biomarkers.