• [2024/05/16] 🔥 Released implementation of LangCell (📃Paper, 💻Code, 🤖Model, 📎Citation).

  LangCell is the first "language-cell" multimodal pre-trained model jointly developed by PharMolix and Institute for AI Industry Research (AIR) . It effectively enhances the understanding of single-cell transcriptomics by learning knowledge-rich texts containing cell identity information, and addresses the task of cell identity understanding in data-scarce scenarios. LangCell is the only single-cell model capable of effective zero-shot cell identity understanding and has also achieved SOTA in few-shot and fine-tuning scenarios. LangCell will soon be integrated into OpenBioMed.

• [2023/08/14] 🔥 Released implementation of BioMedGPT-10B (📃Paper, 🤖Model, 📎Citation), BioMedGPT-LM-7B (🤗HuggingFace Model) and DrugFM (🤖Model).

  BioMedGPT-10B is the first commercial-friendly multimodal biomedical foundation model jointly released by PharMolix and Institute of AI Industry Research (AIR) . It aligns the language of life (molecular structures and protein sequences) with human natural language, performing on par with human experts on biomedical QA benchmarks, and demonstrating powerful performance in cross-modal molecule and protein question answering tasks. BioMedGPT-LM-7B is the first commercial-friendly generative foundation model tailored for biomedicine based on Llama-2.

  DrugFM is a multi-modal molecular foundation model jointly developed by Institute of AI Industry Research (AIR) and Beijing Academy of Artificial Intelligence, BAAI. It leverages UniMAP, a pre-trained molecular model that captures fine-grained properties and representations of molecules, and incorporates MolFM, our multimodal molecular foundation model. DrugFM achieves SOTA on cross-modal retrieval.

• [2023/06/12] Released implementation of MolFM (📃Paper, 🤖Model, 📎Citation) and CellLM (📃Paper, 🤖Model, 📎Citation).

  MolFM is a multi-modal molecular foundation model that enables joint comprehension of molecular structures, biomedical documents and knowledge graphs. On cross-modal retrieval, MolFM outperforms existing models by 12.03% and 5.04% under zero-shot and fine-tuning settings. MolFM also excels in molecule captioning, text-to-molecule generation and molecule property prediction.

  CellLM is the first large-scale cell representation learning model trained on both normal cells and cancer cells with divide-and-conquer contrastive learning. CellLM beats ScBERT on cell type annotation (71.8 vs 68.8), few-shot single-cell drug sensitivity prediction (88.9 vs 80.6) and single-omics cell line drug sensitivity prediction (93.4 vs 87.2).

• [2023/04/23] Released implementation of BioMedGPT-1.6B (🤖Model) and OpenBioMed.

• Introduction
• Installation
• Quick Start
• Documentations
• Limitations
• Cite us

This repository holds OpenBioMed, a Python deep learning toolkit for AI-empowered biomedicine. OpenBioMed provides easy access to multimodal biomedical data, i.e. molecular structures, transcriptomics, knowledge graphs and biomedical texts for molecules, proteins, and single cells. OpenBioMed supports a wide range of downstream applications, ranging from traditional AI drug discovery tasks to newly-emerged multimodal challenges.

OpenBioMed provide researchers with easy-to-use APIs to:

• 3 different modalities for molecules, proteins, and single cells: molecular structures or transcriptomics, biomedical texts, and knowledge graphs. OpenBioMed provide a unified pipeline for researchers to access, process, and fuse these modalities.
• 10 downstream tasks, categorized into AI drug discovery (AIDD) tasks like drug-target binding affinity prediction and molecule property prediction, as well as multimodal tasks like molecule captioning and text-based molecule generation.
• 20+ deep learning models, including BioMedGPT-10B, MolFM, CellLM. Researchers can flexibly compose different components to curate their own models.
• 20+ datasets that are most popular in AI-driven biomedical research. Reproductible benchmarks with abundant model combinations and comprehensive evaluations are provided.

Key features of OpenBioMed include:

• Unified Data Processing Pipeline: easily load and transform the heterogeneous data from different biomedical entities and modalities into a unified format.
• Off-the-shelf Inference: publicly available pre-trained models and inference demos, readily to be transferred to your own data or task.
• Reproducible Model Zoo: flexibly replicate and extend state-of-the-art models on existing and new applications.

The following table shows the supported tasks, datasets and models in OpenBioMed. This is a continuing effort and we are working on further growing the list.

1. (Optional) Creating conda environment:

conda create -n OpenBioMed python=3.9
conda activate OpenBioMed

2. Install required packages:

pip install -r requirements.txt

3. Install Pyg dependencies:

pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-(your_torch_version)+(your_cuda_version).html
pip install torch-geometric
# If you have issues installing the above PyTorch-related packages, instructions at https://pytorch.org/get-started/locally/ and https://github.com/pyg-team/pytorch_geometric may help. You may find it convenient to directly install PyTorch Geometric and its extensions from wheels available at https://data.pyg.org/whl/.

Note: additional packages may be required for some downstream tasks.

Checkout our Jupytor notebooks and documentations for a quick start!

This repository holds BioMedGPT-LM-7B and BioMedGPT-10B, and we emphasize the responsible and ethical use of these model. BioMedGPT should NOT be used to provide services to the general public. Generating any content that violates applicable laws and regulations, such as inciting subversion of state power, endangering national security and interests, propagating terrorism, extremism, ethnic hatred and discrimination, violence, pornography, or false and harmful information, etc. is strictly prohibited. BioMedGPT is not liable for any consequences arising from any content, data, or information provided or published by users.

This repository is licensed under the MIT License. The use of BioMedGPT-LM-7B and BioMedGPT-10B models is accompanied with Acceptable Use Policy.

We are looking forward to user feedback to help us improve our framework. If you have any technical questions or suggestions, please feel free to open an issue. For commercial support or collaboration, please contact [email protected].

If you find our open-sourced code and models helpful to your research, please consider giving this repository a 🌟star and 📎citing the following articles. Thank you for your support!

@misc{OpenBioMed_code,
      author={Luo, Yizhen and Yang, Kai and Hong, Massimo and Liu, Xing Yi and Zhao, Suyuan and Zhang, Jiahuan and Wu, Yushuai and Nie, Zaiqing},
      title={Code of OpenBioMed},
      year={2023},
      howpublished={\url{https://github.com/BioFM/OpenBioMed.git}}
}

@misc{luo2023biomedgpt,
      title={BioMedGPT: Open Multimodal Generative Pre-trained Transformer for BioMedicine}, 
      author={Yizhen Luo and Jiahuan Zhang and Siqi Fan and Kai Yang and Yushuai Wu and Mu Qiao and Zaiqing Nie},
      year={2023},
      eprint={2308.09442},
      archivePrefix={arXiv},
      primaryClass={cs.CE}
}

@misc{luo2023empowering,
      title={Empowering AI drug discovery with explicit and implicit knowledge}, 
      author={Yizhen Luo and Kui Huang and Massimo Hong and Kai Yang and Jiahuan Zhang and Yushuai Wu and Zaiqing Nie},
      year={2023},
      eprint={2305.01523},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@misc{luo2023molfm,
      title={MolFM: A Multimodal Molecular Foundation Model}, 
      author={Yizhen Luo and Kai Yang and Massimo Hong and Xing Yi Liu and Zaiqing Nie},
      year={2023},
      eprint={2307.09484},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM}
}

@misc{zhao2023largescale,
      title={Large-Scale Cell Representation Learning via Divide-and-Conquer Contrastive Learning}, 
      author={Suyuan Zhao and Jiahuan Zhang and Zaiqing Nie},
      year={2023},
      eprint={2306.04371},
      archivePrefix={arXiv},
      primaryClass={cs.CE}
}

@misc{zhao2024langcell,
      title={LangCell: Language-Cell Pre-training for Cell Identity Understanding}, 
      author={Suyuan Zhao and Jiahuan Zhang and Yizhen Luo and Yushuai Wu and Zaiqing Nie},
      year={2024},
      eprint={2405.06708},
      archivePrefix={arXiv},
      primaryClass={q-bio.GN}
}