kuleshov-group/

caduceus-ph_seqlen-131k_d_model-256_n_layer-16

To use the pre-trained model for masked language modeling, use the following snippet:

Alternatively, you can instantiate a model from scratch to train on your own data as follows:

This is the Caduceus-Ph model with hidden dimension 256 and 16 MambaDNA layers. This model is not inherently reverse complement (RC) equivariant. Rather, it was pre-trained using RC data augmentation. Its intended usage is as follows: for downstream tasks, the model should be trained with RC data augmentation. At downstream task inference, the model should be run twice: once on a sequence and once on its RC. The output of these two applications should be combined (averaged) to form the downstream task prediction.

This model was pre-trained on the human reference genome with sequence length 131,072 for 50k steps (each step contained ~1M base pairs / tokens).

For more details, please see our paper: Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling.

Please cite our work using the bibtex below:

@article{schiff2024caduceus,
  title={Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling},
  author={Schiff, Yair and Kao, Chia-Hsiang and Gokaslan, Aaron and Dao, Tri and Gu, Albert and Kuleshov, Volodymyr},
  journal={arXiv preprint arXiv:2403.03234},
  year={2024}
}

Yair Schiff (yzs2@cornell.edu)