Mizuki Molecule and Tsinghua University Intelligent Industry Research Institute (AIR) announced the open source of the world's first commercially available multi-modal biomedical 10 billion parameter large model BioMedGPT-10B , which can be used to improve the efficiency of all aspects of drug research and development, including new drug project evaluation, drug design and optimization, clinical trial design, indication expansion, etc.
In addition, the model's question answering ability in the field of biomedicine is comparable to that of human experts, and it has reached SOTA in natural language, molecular, and protein cross-modal question answering tasks, and has successfully passed the US Physician Qualifying Examination.
Open source address:
BioMedGPT is a brand-new multimodal semantic understanding framework. It uses the pre-trained large language model in the biomedical field—BioMedGPT-LM as a bridge to connect natural language, biological coding language, and chemical molecular language.
BioMedGPT Architecture::
BioMedGPT-LM fine-tunes the general-purpose large-scale language model based on the GPT architecture by making full use of massive biomedical-related data to achieve better performance in the biomedical field.
As a connecting bridge, BioMedGPT-LM can connect codes of various biological modalities, including molecular, protein, cell, and gene expression data, and can also integrate expertise embodied in knowledge graphs, documents, numerical experiment results, and other formats. Through the integration of cross-modal feature fusion modules, different modal biological coding languages, chemical molecular languages and natural languages can be integrated in the same feature space.
At the same time, Mizuki Molecule and AIR jointly open sourced the world's first free commercial Llama 2 language model BioMedGPT-LM-7B dedicated to biomedicine . "AIR-Zhiyuan Health Computing Joint Research Center" cooperated to open source the basic model of small molecule drug DrugFM. The open-source basic model of biomedicine is scientific research-oriented and commercially available, providing a large model base for biomedical research and application.