VantAI, Inc.

A Roivant Sciences spinout says it has made the first AI model, to its knowledge, that simultaneously predicts biomolecule structures while generating new molecules. The diffusion-based model, called Neo-1, can generate new molecular glues, or small molecules that get two proteins to stick together, VantAI’s leaders exclusively told Endpoints News . The biotech will publish a 6,000-word blog post Friday describing the model’s performance and broader potential in drug discovery. All told, VantAI sees Neo-1 as going beyond the capabilities of AlphaFold 3, the state-of-the-art model developed by Google DeepMind and Isomorphic Labs — especially to help make new glues. “It tackles the big next step after AlphaFold 3,” VantAI’s chief technology officer Luca Naef said in an interview. “What we’re really excited about is unlocking the next chapter of this modality, so not just degradation but going beyond.” Neo-1’s unveiling is the most substantive update yet on what VantAI has been working on since its 2019 founding. The startup has previously announced research partnerships to discover new molecular glues with Johnson & Johnson, Bristol Myers Squibb and Blueprint Medicines. Neo-1 could allow VantAI to stand out in an increasingly crowded glue space , since it can generate new candidates from scratch. Roivant and the Korean conglomerate SK Group hold equity in VantAI, according to regulatory filings. VantAI now has about 50 employees with offices in New York and Zurich, as well as a data-generation facility in Berlin, CEO Zachary Carpenter said. Neo-1’s release figures as a key part of VantAI’s pitch to potential investors. The ability to simultaneously predict structure while generating a new molecule may crack a unique, thorny problem related to finding glues. Glues work by linking two proteins in a cell, creating a biomolecular complex — the two proteins and the linking glue molecule — that doesn’t otherwise form. Today’s leading AI models, such as AlphaFold 3 and similar open-source versions like Boltz-1, struggle with these complicated structures, particularly because there are so few examples of them in datasets like the Protein Data Bank. Neo-1, by contrast, can take a mix of sequences and structures of biomolecules as an input to predict the overall structure and generate a new molecule that fits that prediction. “It’s impossible to generate something if you don’t know the molecule that will form it,” Carpenter said. “Having a model that can take two sequences as input and generate that molecule unlocks a lot of potential in this space, and that’s exactly why this model was built.” VantAI is betting that Neo-1 can deliver a whole slew of new glues. The broader field uses the same group of 10 or 15 glues, Carpenter said, because it’s been so difficult to discover or create new ones. VantAI’s own biological dataset, called NeoLink, was key in training the model. The idea somewhat mirrors shotgun sequencing, where DNA fragments are reassembled into a picture of the genome. In this case, VantAI tested massive numbers of chemical linkers and proteins, creating a training dataset of its own fragments of proteins and molecules that only holds value for AI models, rather than human researchers. “Five years ago, this data was rather useless,” Naef said. “Now we have folding models where we can put this as input and use it to reconstruct the complex in cases where this is not possible without any data.” Michael Bronstein, a leading figure in AI bio who is also chief scientist-in-residence of VantAI, called NeoLink a “very good example of blackbox data.” Bronstein and Naef wrote a 10,000-plus word treatise last year on the future of blackbox data, or datasets that are not interpretable or valuable to human eyes. A dataset like NeoLink “makes sense only in conjunction with the appropriate machine-learning model,” Bronstein said. For now, Neo-1 and its results are being released only in a technical blog post. Carpenter and Naef said a richer technical paper is on the horizon. They are also considering releasing a portion of the NeoLink dataset for academic research. The blog post acknowledges that the team described retrospective tests of the model, with plans to share prospective and experimentally validated results in the future.

With much of its growth aspirations resting on its flagship oncology portfolio, AstraZeneca set its sights on a new early-stage programme from Pinetree Therapeutics on Wednesday. The deal gives the UK pharma an exclusive option to license an EGFR-targeted protein degrader currently in preclinical development for drug-resistant tumours. Pinetree will receive $45 million in upfront and near-term payments and is also eligible for more than $500 million in milestones, plus tiered royalties. The candidate was developed using the biotech’s AbReptor platform, which enables the design of modular multi-specific antibodies that engage with an effector protein to degrade a membrane-bound, extracellular protein of interest. According to Pinetree CEO Hojuhn Song, its pan-EGFR programme has “demonstrated promising preclinical anti-tumor activity in drug- and TKI-resistant tumours as well as enhanced activity when used in combination with current EGFR inhibitors."Degrader trendsThe deal with Pinetree comes as AstraZeneca aims to boost its annual sales to $80 billion by 2030 as it pursues a “new era of growth” for its oncology, biopharmaceuticals and rare disease portfolio.About half of that sales target is expected to come from its cancer products, which currently contribute about 40% to its annual revenue. For related analysis, see Vital Signs: Where AstraZeneca can find $35 billion by 2030.AstraZeneca is the latest large pharma to secure a targeted protein degradation collaboration this year, albeit with a significantly lower total deal value.Novo Nordisk entered the space this year in a $1.46-billion tie-up with Neomorph, and Novartis agreed to a potentially billion-dollar-plus licensing deal with Arvinas.Bristol Myers Squibb has the most deals in the space, closing six since 2022. Most recently, the pharma partnered with VantAI to use the latter’s generative AI platform to design molecular glues. For related deal analysis, see Vital Signs: Mapping protein degradation partnerships.According to AstraZeneca’s website, it has an automated, end-to-end, orally bioavailable protein degrader discovery platform, as well as a suite of pharmacology and safety assays to measure the level of degradation and predict disease efficacy. The drugmaker also has several preclinical degrader tool compounds available for partnering through its Open Innovation research programme. Last year, AstraZeneca teamed up with the Francis Crick Institute and Imperial College London to discover new molecular glue degraders.

Takeda joined forces with Degron Therapeutics to develop molecular glue degraders for multiple targets in oncology, neuroscience and inflammation. The deal, announced Thursday, is potentially worth up to $1.2 billion, including an undisclosed upfront payment from the Japanese firm.The companies will utilise Degron's GlueXplorer platform to identify, validate and optimise molecular glue degraders for specific therapeutic targets selected by Takeda. Upon reaching a certain stage of advancement, the projects would be transitioned from the Chinese pharma to Takeda for further development and commercialisation.Chris Arendt, Takeda’s head of research, noted that the collaboration “not only adds an innovative new platform to our drug discovery toolbox, it is also an example of cutting-edge innovation emerging in the exciting China biotech sector.” Founded in 2021, Degron raised $22 million via a Series A financing the following year to fund its own pipeline, which includes five oncology programmes.Along with the initial payment, Degron is eligible for milestones and tiered sales royalties on products stemming from the collaboration, which can also be expanded to include further targets. Meanwhile, Takeda will make an equity investment in Degron. Molecular glue degraders continue to garner interest, with a number of recent deals in the space. In February, Bristol Myers Squibb agreed to use VantAI's generative AI platform to design molecular glues, while Novo Nordisk entered the field via a deal with Neomorph. For related analysis, see Spotlight On: BMS leads big pharma in protein degradation deals.ASCO Daily Digest – your go-to-source for the key developments emerging from this year's American Society of Clinical Oncology (ASCO) annual meeting. Exclusively for PLUS subscribers – sign up here!