VantAI has officially announced the launch of its Neo-1, which happens to be the world’s most advanced atomistic foundation model, as well as the first to unify de novo molecular generation with multimodal structure prediction.
According to certain reports, the stated model arrives bearing an ability to make the systematic re-wiring of protein interactions possible, something it does for the purpose of empowering drug developers to precision-engineer novel therapeutics, including molecular glues, bifunctional degraders, and other proximity-based modalities.
More on the same would reveal how Neo-1 brings to the fore a novel approach for the field. We get to say so because a majority of those traditional all-atom structure foundation models do little more than just predicting atomic coordinates. VantAI’s latest brainchild, on the other hand, can effectively generate latent representations of molecules, something which can then be decoded into their three-dimensional structure.
Such a methodology makes it possible for Neo-1 to predict the structure of protein-small molecule complexes and generate entirely new molecules tailor made for a given pocket. Not just that, by doing so, the model conceives fresher avenues for drug design, including for novel molecular glues and other emerging proximity modulation therapeutics.
“Our vision is to make protein interactions programmable,” said Luca Naef, Co-Founder & CTO of VantAI who led the development of Neo-1. “This is only possible if we can precision-design the structure of novel biomolecules, including drug-like small molecules. Existing models allow us to decode the structures of molecules, but to make new medicines we need to be able to design them. Neo-1, for the first time, allows for this in a single unified model.”
Talk about the whole value proposition on a slightly deeper level, we begin from its promise to provide multimodal inputs. These inputs, on their part, will tread up a long distance to help users accept any combination of (partial) sequence, (partial) structure, and experimental data, such as NeoLink-derived restraints or almost any real-world empirical constraint.
Next up, there is the potential for unified de novo molecule generation and structure prediction, which involves supporting atomic-resolution design and reconstruction of protein interfaces, critical for challenging mechanisms of action (MOAs) like molecular glues.
“Neo-1 is a breakthrough model that enables all-atom de novo design across biomolecular space in addition to remarkable multimodal structure prediction capabilities,” said Anthony Costa, Director of Digital Biology at NVIDIA. “Acceleration of Neo-1 on the NVIDIA platform is crucial to accelerate scientific advances in the hands of scientists using the model.”
Another detail worth a mention is rooted in a facility dedicated towards facilitating massive-scale training. This translates to how the solution in question is among the largest diffusion-based models throughout biology, trained on structural and synthetic datasets using hundreds of NVIDIA H100 GPUs.
Rounding up highlights would be the availability of custom datasets and tools. Here, the new model packs together PINDER & PLINDER (co-developed with NVIDIA), and synthetic data created using GPU-mmseqs2, a tool developed by NVIDIA and SNU, slated for deployment via NVIDIA’s recently announced MSA-Search NIM.
We also ought to mention how Neo-1 has been trained using VantAI’s proprietary NeoLink dataset. This access to a vast new type of data and broad programmability should lead to high performance, enabling precise molecular design and reconstruction of protein interface structures at atomic resolution. It also unlocks joint capabilities that are critical for conceiving de novo design for challenging MOAs such as molecular glues.
All in all, VantAI will bank upon the given mechanism to generate active small molecules for previously undruggable targets in its pipeline, and it will do that in a time frame of weeks rather than years.
“Neo-1 unlocks the potential of VantAI’s high-throughput ‘black-box’ data platform, NeoLink. VantAI’s unique approach to data and AI will allow us to illuminate the geometric vocabulary of protein interaction at the full proteome scale, a step necessary to design next-generation therapeutics currently beyond reach. Being both one of the leading frontier AI labs in the bio space and having such a differentiated data source makes VantAI truly unique,” said Michael Bronstein, Chief Scientist at VantAI and DeepMind Professor of AI at the University of Oxford.
