eccDNAMamba/eccDNAMamba_eccdna_copy_number_level_prediction_threshold_4.5

eccDNAMamba

A Pre-Trained Model for Ultra-Long eccDNA Sequence Analysis

---

Model Overview

eccDNAMamba is a bidirectional state-space model (SSM) designed for efficient and topology-aware modeling of extrachromosomal circular DNA (eccDNA).
By combining forward and reverse Mamba-2 encoders, motif-level Byte Pair Encoding (BPE), and a lightweight head–tail circular augmentation, it captures wrap-around dependencies in ultra-long (10–200 kbp) genomic sequences while maintaining linear-time scalability.
The model provides strong performance across cancer-associated eccDNA prediction, copy-number level estimation, and real vs. pseudo-eccDNA discrimination tasks.

---

Quick Start

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("eccdna/eccDNAMamba-1M")
model = AutoModelForMaskedLM.from_pretrained("eccdna/eccDNAMamba-1M")

sequence = "ATGCGTACGTTAGCGTACGT"
inputs = tokenizer(sequence, return_tensors="pt")
outputs = model(**inputs)

# Access logits or reconstruct masked spans
logits = outputs.logits

---

Citation

@inproceedings{
liu2025eccdnamamba,
title={ecc{DNAM}amba: A Pre-Trained Model for Ultra-Long ecc{DNA} Sequence Analysis},
author={Zhenke Liu and Jien Li and Ziqi Zhang},
booktitle={ICML 2025 Generative AI and Biology (GenBio) Workshop},
year={2025},
url={https://openreview.net/forum?id=56xKN7KJjy}
}