Neo-1 - VantAI's first atomic generative AI model
Neo-1, launched by VantAI, is the world's first AI model to unify de novo molecule generation and atomic-level structure prediction. It can predict biomolecular structures, generate entirely new molecules, and has shown promise in designing novel therapeutics such as molecular gels...
What is Neo-1?
Neo-1, launched by VantAI, is the world's first AI model to unify de novo molecule generation and atomic-level structure prediction. It can predict biomolecular structures, generate novel molecules, and excel in designing novel therapeutic drugs such as molecular gels. Neo-1's multimodal input capability can accept various information, including partial sequences, partial structures, and experimental data, significantly improving the efficiency and accuracy of drug design. Combined with VantAI's NeoLink platform, Neo-1 generates sparse structural constraints through cross-linking mass spectrometry and then assembles them into complete atomic-resolution structures, advancing the development of structural biology.
Neo-1's main functions
- Unified Generation and PredictionNeo-1 is the first model to unify in-situ molecule generation with atomic-level structure prediction. By generating latent representations of molecules, rather than predicting atomic coordinates, it can predict biomolecular structures and generate entirely new molecules.
- Multimodal inputNeo-1 accepts inputs from multiple modalities, including partial sequences, partial structures, and experimental data. This multimodal input approach significantly improves the model's flexibility and applicability.
- Large-scale trainingNeo-1 is one of the largest diffusion-based models in biology, trained on structural and synthetic datasets using hundreds of NVIDIA H100 GPUs.
- Custom datasets and toolsNeo-1 combines VantAI’s proprietary NeoLink dataset with the PINDER & PLINDER tools developed in collaboration with NVIDIA.
The technical principles of Neo-1
- Diffusion processes in potential spaceNeo-1 shifts the diffusion process from traditional coordinate space to a latent space. This shift allows the model to reason within a smoother sequence and structural landscape, generating entirely new molecules, including proteins, peptides, and small molecules, while predicting their structures with atomic-level precision.
- Large-scale training and custom datasetsNeo-1 is one of the largest diffusion-based models in biology, trained on structural and synthetic datasets using hundreds of NVIDIA H100 GPUs. It leverages VantAI's proprietary NeoLink dataset and the PINDER & PLINDER tools co-developed with NVIDIA to enhance model performance.
- Precise molecular generation and structure predictionNeo-1 employs a coarse-to-fine generative approach, applying intermediate rewards based on the entire molecular structure to guide molecular generation towards any goal. This differs from traditional autoregressive models, which lack flexibility in the generation process.
Neo-1 project address
- Project official website:https://www.vant.ai/neo-1
Application scenarios of Neo-1
- Molecular adhesive designNeo-1 can design novel therapeutics such as molecular gels targeting complex targets, reducing the time that traditionally takes years to weeks.
- Protein complex structure predictionNeo-1 can predict the structure of a variety of biomolecular complexes, including ternary complexes, antibody-antigen interactions, and protein-peptide complexes.
- Applications of the NeoLink data platformNeo-1, combined with VantAI’s NeoLink data platform, can assemble complete atomic-resolution structures based on sparse structure constraints generated by cross-linking mass spectrometry.
- Antibody discoveryNeo-1 enables rational antibody discovery from start to finish. It can take partial antibody sequences and antigen structures as input, and simultaneously fold VH antibody fragments to generate partial CDRH3 sequences.