Chai-1 is a molecular structure prediction foundation model released by Chai Discovery in the fall of 2024. It is a next-generation AI-based bio tool that predicts the 3D structure of various biomolecular complexes, including not only proteins but also small molecules, DNA, and RNA, with high accuracy. Just as a large language model (LLM) understands the meaning and relationships of words based on context and generates text, Chai-1 precisely synthesizes the input amino acid and nucleotide sequences with molecular graph structure information to derive 3D spatial coordinates at the individual atom level. Within a single model environment, it can simultaneously model multi-modal data, including protein-protein interactions and protein-ligand binding structures, making it a core framework for discovering new drug candidates and conducting biomolecular design research. Existing molecular structure prediction systems, such as AlphaFold2 and ESMFold, had limitations in that the MSA (Multiple Sequence Alignment) pre-processing step, which involves searching a large biological database, consumed a significant amount of time and computational resources, ranging from several minutes to tens of minutes. AlphaFold3, a proprietary model, demonstrated high prediction performance, but its academic and commercial use was restricted, making it difficult to apply and integrate into custom pipelines in research settings. In contrast, Chai-1 provides powerful, high-precision 3D structure prediction performance with only a single sequence input, completing inference in just a few seconds without the MSA generation process. It also makes the entire code and model weights openly available under the Apache 2.0 license, allowing for free use in commercial research and development without restrictions. Furthermore, it incorporates the flexibility to directly input epitope binding sites or spectroscopic constraints obtained from wet-lab experiments as pipeline-driving prompts, dramatically improving prediction accuracy. In actual biotechnology and pharmaceutical research settings, Chai-1 is widely used for predicting Binding Affinity and conformational binding poses between target proteins and organic small molecule compounds. Researchers can set the target protein sequence and the SMILES structure of a low-molecular-weight drug candidate as input, and by calling the Chai-1 Python API on a high-performance GPU, they can derive the 3D structure of the drug-target complex, reducing the virtual screening speed by more than 10 times compared to previous methods. In addition, in the field of antibody design, researchers can simultaneously input the sequence of the desired antigen site and specific epitope residue constraints to perform ultra-fast simulations of antibody-antigen binding structures, and based on the calculated RMSD (Root Mean Square Deviation) and LDDT (Local Distance Difference Test) scores, they can build a pipeline to efficiently screen for effective binding candidates. As a result, Chai-1 is expanding structural biology research from simulation-based virtual exploration to the realm of prediction and perfectly controllable design. From multi-chain complex structure analysis to high-molecular RNA structure simulation, researchers can generate structural biology datasets and perform structure optimization with just a few lines of Python code. This open-source structural foundation model will drive innovation in the global life science research ecosystem as a core engine for automated drug development platforms and on-premise AI drug screening pipelines in the future.

Chai-1 is a molecular structure prediction foundation model released by Chai Discovery in the fall of 2024. It is a next-generation AI-based bio tool that predicts the 3D structure of various biomolecular complexes, including not only proteins but also small molecules, DNA, and RNA, with high accuracy. Just as a large language model (LLM) understands the meaning and relationships of words based on context and generates text, Chai-1 precisely synthesizes the input amino acid and nucleotide sequences and molecular graph structure information to create a 3D structure at the individual atom level.

Case 1: Predicting the 3D conformation of a target protein and small molecule ligand complex, and performing virtual screening for drug binding poses.