CARMANIA

Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis

CARMANIA is a self-supervised genomic language model framework that augments next-token prediction with a transition-matrix regularization loss. This integration improves biological sequence modeling by aligning predicted transitions with empirical bigram(2-mer) statistics, allowing for better long-range dependency modeling and functional interpretation.

---

🧠 Pretrained Models

The following models are already available for use on Hugging Face Hub:

• 🦠🧬 MsAlEhR/carmania-big-10k-prok-genome
• 🦠🧬 MsAlEhR/carmania-4k-scp-gene-taxa
• 👤🧬 MsAlEhR/carmania-160k-seqlen-human

---

🚀 Quick Start

from transformers import AutoModel, AutoTokenizer
import torch

model = AutoModel.from_pretrained(
    "MsAlEhR/carmania-160k-seqlen-human",
    trust_remote_code=True,
    torch_dtype=torch.float16,   # fixed dtype (or autocast)
).to("cuda")

tokenizer = AutoTokenizer.from_pretrained(
    "MsAlEhR/carmania-160k-seqlen-human",
    trust_remote_code=True,
    model_max_length=160000,
)

inputs = tokenizer("ACGTAGGCTA", return_tensors="pt").to("cuda")

outputs = model(**inputs)

🧪 Sequence-Guided Generation

An experimental notebook exploring CARMANIA-driven sequence optimization using Enformer scores is now available.
This lightweight module perturbs input DNA sequences and uses Enformer’s predicted regulatory signals as a scoring function to iteratively generate variants with improved activity.

📄 Notebook:
carmania_enformer_guided_generation.ipynb

Citation

@article{refahi2026context,
  title={Context-aware regularization with markovian integration for attention-based nucleotide analysis},
  author={Refahi, Mohammad Saleh and Abavisani, Mahdi and Sokhansanj, Bahrad and Brown, James R and Rosen, Gail},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  pages={108518--108544},
  year={2026}
}