eccDNA2Ca: an ensemble deep learning framework for interpretable prediction of cancer-associated extrachromosomal...

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Extrachromosomal circular DNA (eccDNA) drives oncogene amplification, therapeutic resistance, and cancer heterogeneity.
• Current computational methods for eccDNA lack focus on functional and clinical significance in oncogenesis.

Purpose of the Study:

• To develop eccDNA2Ca, an interpretable framework for predicting cancer-associated eccDNAs from raw sequences.
• To enable functional studies and biomarker discovery by prioritizing eccDNAs.

Main Methods:

• Developed eccDNA2Ca, an ensemble learning framework combining XGBoost, CNN, and LSTM.
• Trained models on a curated dataset of 465 human eccDNAs across 16 cancer types.
• Utilized genomic features, deep sequence encodings, SHAP analysis, and motif discovery for interpretability.

Main Results:

• eccDNA2Ca achieved superior performance (AUC > 0.96, AUPR > 0.95) and outperformed conventional classifiers.
• Feature interpretation highlighted repeat content and k-mer frequencies; motif analysis found transcription factor binding sites.
• Validated associations between predicted eccDNAs and tumor mutational burden, immune infiltration, microsatellite instability, and patient survival.

Conclusions:

• eccDNA2Ca is the first interpretable, open-source tool to prioritize cancer-relevant eccDNAs based on sequence features.
• Provides a scalable resource for eccDNA functional studies, cancer biomarker discovery, and translational genomics.
• Source code and web server are publicly available for broader research application.