HyenaCircle: a HyenaDNA-based pretrained large language model for long eccDNA prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Extrachromosomal circular DNA (eccDNA) are circular DNA molecules derived from chromosomes with significant roles in various diseases.
• Detecting long eccDNAs (1-5 kb) is challenging due to their size, limiting functional studies.
• Existing methods struggle with the accurate identification and prediction of long eccDNAs.

Purpose of the Study:

• To introduce HyenaCircle, a novel deep learning model for predicting the formation of long eccDNAs.
• To leverage third-generation sequencing data and large language models for enhanced eccDNA detection.
• To overcome the limitations of current methods in identifying large eccDNAs.

Main Methods:

• HyenaCircle was developed by adapting the HyenaDNA model with a classifier head, utilizing Nanopore sequencing data and flanking sequences.
• Data augmentation, regularization, and class imbalance weighting were employed to enhance model robustness.
• The FLED algorithm identified full-length eccDNAs, which were then extended and paired with negative controls.

Main Results:

• HyenaCircle achieved a validation AUROC of 0.715 and recall of 0.776, outperforming DNABERT by 5.9% in AUROC.
• Ablation studies confirmed the importance of flanking sequences for model stability.
• The model demonstrated superior stability and convergence compared to the baseline HyenaDNA architecture.

Conclusions:

• HyenaCircle effectively integrates third-generation sequencing and large language models for accurate long eccDNA prediction.
• The HyenaDNA architecture proves adept at long-sequence genomic modeling, offering new insights into eccDNA identification.
• This work presents a significant advancement in the prediction and characterization of disease-associated eccDNAs.