This is the official code repo of the paper: UNICORN: Towards Universal Cellular Expression Prediction with an Explainable Multi-Task Learning Framework.

To install UNICORN, please refer to the content in the yml file "unicorn.yml". It is modified from the packages required by seq2cells.

If you are interested in individualized gene expression prediction, you may need to install bcftools and samtools.

conda install bcftools
conda install samtools

These instructions can help you prepare all the required files for running UNICORN.

Please check the folder seq2cells for details of the model structure. The files under the models contain the variations used in UNICORN.

If you intend to use different DNA sequence embeddings for predicting gene expression, please refer BEND for details.

If you intend to use gene embeddings for natural language, please refer scELMo for details.

We provide a tutorial for training and evaluating UNICORN in the tutorial folder. To visualize the prediction results of peaks, please install pygenometracks with:

pip install pygenometracks==3.8

Please check this link as well as this link for accessing the datasets used for training.

To use GTEx data, please visit this website to request access to protected data.

We answer some general questions in this section:

1. Our training code is summarized in this folder. As we develop our methods based on seq2cells, we modify their methods and loss functions to support our requirements.

2. If your data have multi-modality, you want to choose modality-specific embeddings. For example, to predict CITE-seq data, your input embeddings could be both Enformer embeddings (for gene expression) and ESM embeddings (for protein expression).

3. Please refer the tutorial section to learn about model training, and multi-modal as well as single-modal share the same training pipeline.

4. If you have any questions when running our codes/reproducing our results, feel free to leave an issue in this repo😄！

We thank the authors of seq2cells, bcftools, and samtools for their great codes.

@article{liu2025unicorn,
  title={UNICORN: Towards universal cellular expression prediction with a multi-task learning framework},
  author={Liu, Tianyu and Huang, Tinglin and Wang, Lijun and Lin, Yingxin and Ying, Rex and Zhao, Hongyu},
  journal={Nature Communications},
  volume={16},
  number={1},
  pages={9455},
  year={2025},
  publisher={Nature Publishing Group UK London}
}