[ICLR2025] ConvNova 🧬 Revisiting Convolution Architecture in the Realm of DNA Foundation Models

OpenReview |  arXiv |  GitHub |  HuggingFace 🤗(coming soon)

ConvNova demonstrates that, if carefully designed, a pure CNN can serve as a DNA foundation model that surpasses Transformer and SSM-inspired architectures, while retaining the classic convolutional advantages of stronger locality bias, lower memory footprint, and markedly faster training and inference.

---

🚩 Plan

• Scripts for Pretraining, NT & Genomic Benchmarks.
• Paper Released.
• Pretrained Weights of ConvNova.
• Source Code and Pretrained Weights on transformers.
• Scripts for DeepSEA & Bend-gene-finding.

---

1 Quick start

Clone the repo.

  git clone [email protected]:aim-uofa/ConvNova.git
  cd ConvNova/convnova

Prepare conda env.

  conda create -n convnova python==3.10
  conda activate convnova
  pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 
  pip install -r requirements.txt --no-deps
  pip install pytorch-lightning==1.8.6 --no-deps
  pip install packaging --no-deps

  pip install lightning_utilities --no-deps
  pip install torchmetrics
  pip install tensorboardX

Download the data.(Pretrain)

  mkdir data
  mkdir -p data/hg38/
  curl https://storage.googleapis.com/basenji_barnyard2/hg38.ml.fa.gz > data/hg38/hg38.ml.fa.gz
  gunzip data/hg38/hg38.ml.fa.gz  # unzip the fasta file
  curl https://storage.googleapis.com/basenji_barnyard2/sequences_human.bed > data/hg38/human-sequences.bed

You can check out the Nucleotide Transformer ang Genomic Benchmarks paper for how to download and process NT benchmark & Genomic Benchmark datasets.

The final file structure (data directory) should look like

  |____bert_hg38
| |____hg38.ml.fa
| |____hg38.ml.fa.fai
| |____human-sequences.bed
|____nucleotide_transformer
| |____H3K36me3
| |____......
|____genomic_benchmark
| |____dummy_mouse_enhancers_ensembl
| |____....

---

2 Using ConvNova with 🤗 Transformers

Coming Soon

---

3 Reproducing the paper

3.1 Pre-training on the Human Reference Genome

  python train.py experiment='hg38-pretrain/convnova'

you can adjust the hyperparameters by using cmd like following, detailed hyperparameters setting can be seen in configs/experiment/xxx/xxx.yaml

  python train.py experiment='hg38-pretrain/convnova' wandb=null trainer.devices=4

3.2 Genomic Benchmarks (short-range)

GenomicBenchmarks provides 8 binary- and multi-class tasks packaged as a Python library.

Remeber to adjust the setting for different dataset like max seq length and the pretrained checkpoint(comming soon).

  python train.py experiment='genomic-benchmark/convnova' with-some-argments

3.3 Nucleotide Transformer Benchmark

Datasets are hosted on the Hub as InstaDeepAI/nucleotide_transformer_downstream_tasks.

Remeber to adjust the setting for different dataset like max seq length and the pretrained checkpoint(comming soon).

  python train.py experiment='nt-benchmark/convnova' with-some-argments

---

4 Citation

@inproceedings{bo2025convnova,
  title     = {Revisiting Convolution Architecture in the Realm of DNA Foundation Models},
  author    = {Yu Bo and Weian Mao and Yanjun Shao and Weiqiang Bai and Peng Ye
               and Xinzhu Ma and Junbo Zhao and Hao Chen and Chunhua Shen},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2025}
}

---

5 Acknowledgements

ConvNova builds on the training, logging and data-loading scaffolds of HyenaDNA and Caduceus, and evaluates on Genomic Benchmarks, Nucleotide Transformer tasks, and the Long-Range Benchmark. We thank the maintainers of these open resources for making rigorous comparison possible.