Nabla Bio, Inc.

Nearly two years ago, Isomorphic Labs dropped AlphaFold 3, the latest generation of its Nobel Prize-winning , structure-predicting AI model. Now the company claims it has another major advance on its hands. While it hasn’t released extensive details, Tuesday’s news could put Isomorphic significantly ahead of the rival startups and researchers who have built their own clones and advancements on AlphaFold 3 in recent years. In a blog post and technical paper , the London-based startup detailed the performance of what it calls the Isomorphic Labs Drug Design Engine, or IsoDDE. Despite the clunky name, the company claims it has more than doubled the success rate for accurately predicting protein-drug and antibody-antigen structures, compared with AlphaFold 3. The company called the new engine a “step-change improvement” over AlphaFold 3. “These sorts of capabilities are quite profound for us,” Isomorphic’s president Max Jaderberg said in an interview with Endpoints News . “They literally unlock new targets, new potential solutions to patients.” Tuesday’s report is full of results on a range of benchmarks, including for biomolecular structure predictions, binding affinity predictions, and discovering new pockets on target proteins. But the paper is also light on certain technical details, multiple scientists and industry executives outside of Isomorphic said. That includes specifics on what improvements — be it training data, new AI model designs, or something else — drove the gains. In an interview with Endpoints News , Isomorphic’s director of machine learning Michael Schaarschmidt credited the gains to a “combination of everything.” That included major changes to the underlying model, he said, without divulging specifics. “There’s some real changes to the architecture,” Schaarschmidt said. “So it’s a lot of small wins, but also some real big changes.” These types of claims, with few scientific and technical details, have led to criticisms of AI companies conducting “science by press release,” rather than releasing full, peer-reviewed research papers that can be scrutinized by others. Pat Walters, a longtime computational biotech veteran, said the lack of certain methodological details makes it difficult to interpret some of the results. Often, when researchers build models, they use some data to train the model while reserving some data to use as a test. Because of the complexity and size of these datasets, there’s always the risk that data can “leak” between those two groups, possibly letting the model cheat on the exam. “It’s very easy to get data leakage when training an activity prediction model,” Walters said in an email. “The Iso team doesn’t describe how they did their train/test splits, so it’s tough to gauge performance.” But the secrecy trend has become increasingly common as AI startups have grown more competitive and more protective of commercially valuable research. AlphaFold2, for example, was fully open-sourced, while AlphaFold 3 came with more restrictions on its usage. Now with IsoDDE, it’s not clear what models even comprise its drug design engine, making it difficult for independent researchers or competing labs to vet and compare performance. Isomorphic isn’t alone in that. Chai Discovery’s Chai-2, Nabla Bio’s JAM, and Iambic’s NeuralPLexer3 are among many recent models that have been kept proprietary. That makes it harder to directly compare their performance. Despite the caveats, the progress seems to place Isomorphic at the top of the AI biotech world. It has been a leader in the field since its 2021 founding, led by CEO Demis Hassabis, who is also CEO of Google DeepMind. The company has raised $600 million, grown past 300 employees, and polarized the field with grandiose goals of solving all diseases . The breadth of Tuesday’s performance gains — across structures, affinities, and novel protein pockets — is early evidence of traction in Hassabis’ ambition to upend the industry with an AI-first approach to drug design. In prior interviews, Hassabis has described much of the AI bio field as building models specific to an individual task or particular target. His goal has been to see if AI models can generalize, or make accurate predictions on scenarios that fall outside of training data. Tuesday’s report shows progress on generalization, which has largely evaded the field. On protein-molecule structures, Isomorphic says it achieved a 50% success rate on a category of difficult programs, specifically designed to test if AI models are actually learning biology, or just memorizing from their training data. AlphaFold 3, for instance, scored 23% on that metric, which still beats models like Boltz-1 and Chai-1. Isomorphic’s newest version more than doubled the AlphaFold 3 performance. James Fraser, a bioengineering professor at the University of California, San Francisco, acknowledged the results as impressive, saying they were “something that has seemed far beyond other groups.” “It will likely take some time (and likely hints of ‘how’ it improved) for the field to catch up,” Fraser wrote in an email. Much of the technical paper directly compares Isomorphic’s latest model with Boltz-2, a popular open-source protein model . A new startup, called Boltz, launched last month with $28 million from investors like a16z and Amplify Partners to sell software to life sciences customers around models like Boltz-2. IsoDDE was more than twice as good as AlphaFold 3 in predicting antibody-antigen structures. And it was about 20 times better than Boltz-2 on that task, according to Tuesday’s release. Boltz CEO Gabriele Corso called the results “definitively impressive” in an email, adding his team has also seen further predictions improvements beyond Boltz-2, which was released last June. A driver in starting Boltz, which is headquartered across the street from Isomorphic in London, was the desire to keep leading AI models open and accessible to scientists. So far, Isomorphic has signed partnership deals with a handful of large drugmakers like Eli Lilly, Novartis, and Johnson & Johnson. “The rest of the field will also have access to a significantly better model soon!” Corso wrote. Isomorphic also said its AI models meet or beat an industry gold standard of physics-based simulations. Ramy Farid, the CEO of Schrödinger, a longtime seller of physics-based software, said Isomorphic’s report lacks enough detail and transparency behind how they reported those results, which “could easily lead to incorrect conclusions.” Ultimately, the companies will still have to prove that all of their work can lead to new drugs being tested in patients. Isomorphic is “pushing really hard toward the clinic,” Jaderberg said, without any delays. Hassabis has recently said he expects to start clinical trials by the end of 2026, an apparent pushback to a previous timeline by the end of 2025. (Hassabis previously told Endpoints he had misspoken on that prior timeline, referring to when Isomorphic would select its first preclinical drug candidate.)

Takeda Pharmaceuticals\' project with Iambic Therapeutics will initially focus on the oncology, gastrointestinal and inflammation treatment areas.\n After sharing word that it would be prioritizing artificial intelligence to aid discovery work, Takeda Pharmaceuticals has inked a multiyear deal to access Iambic Therapeutics\' broad suite of drug discovery platforms for up to $1.7 billion biobucks.\"As you think about the AI space, work your way through landscape, and categorize Iambic as company, it is not about AI hype,\" Iambic CEO Tom Miller told Fierce Biotech. \"We are showing in patients that these drugs are better, creating validation.\"Tokyo-based Takeda has decided to leverage Iambic’s software and drug discovery assets to advance certain small molecule programs, with the work starting in the oncology, gastrointestinal and inflammation therapeutic areas. Iambic’s generative model NeuralPLexer, which is designed to predict protein ligand complexes, is also part of the deal. Upfront, research cost and tech access payments are part of the deal, though the companies didn\'t reveal specific numbers. In addition, Iambic could receive milestone-based payments that could exceed $1.7 billion, according to the Feb. 9 press release.The deal\'s financial potential makes it one of the largest AI drug discovery deals to date in a rapidly growing field. The collaboration also builds on Iambic’s other partnerships, a roster that includes Revolution Medicines, Jazz Pharmaceuticals and Lundbeck.“Our collaboration with Takeda is a powerful opportunity to apply our AI-driven discovery and development platform, and we are excited to partner with their team to quickly advance new and better drug candidates,” Miller said in the release. “This collaboration further validates our industry-leading technology and highlights both the breadth of our discovery capabilities and the scale at which we can operate.”  The deal is unique for Iambic because it licenses both software and drug discovery platforms, which support a “rapid design-make-test-analyze” cycle that can accelerate drug development, according to the company. By identifying novel chemical modalities for difficult-to-address targets, Iambic’s platform is built to find diverse candidates for development. \"The Takeda deal is a combination of both tech deployment and involved drug discovery,\" Miller said in the interview. \"It continues traction in the ways which we can monetize the platform.\"Though Iambic is happy to welcome partners, Miller emphasized that the company only collaborates when it feels a potential agreement will make its own pipeline stronger. He says Iambic’s goal is not to create as many partnerships as possible, but to grow its wholly owned pipeline, gain expertise and contribute to scaling its solutions. AI drug discovery biotechs have proliferated, but there are few that have created assets that have made it to the clinical stage and even fewer that have created candidates with clinically differentiated data. Iambic’s most advanced asset, oral oncology candidate IAM1363, delivered promising early-phase clinical data last fall prior to the company’s $100 million fundraise in November.Iambic’s platform is disease agnostic; the company\'s internal efforts and its partnerships have targeted conditions from migraines to cancer as well as gastrointestinal and inflammation-related diseases.For the former Cal Tech professor turned Iambic CEO, the Takeda partnership is further evidence that the firm has moved beyond the AI hype that is so prevalent in the industry. “This deal is reinforcing our thesis that better tech leads to better meds,” Miller said. “It provides a validation and value assignment, showing patients that these drugs are better.” Along with Takeda, other large pharmas and biotechs are looking to strengthen their R&D rosters with AI partners. Last week, Qilu Pharmaceutical locked in a $120 million deal for Insilico Medicine’s AI drug development capabilities in the cardiometabolic disease area.The Iambic deal is more evidence of Takeda’s 2024 strategy shift to focus from 10 modalities to four: small molecules, biologics, antibody-drug conjugates and allogeneic cell therapies. Last fall, the focus shrunk again when Takeda abandoned cell therapy altogether, resulting in 140 layoffs. Takeda’s refined focus resulted in another AI partnership in October with drug designer Nabla Bio and a $1.2 payment to Innovent Biologics for the rights to two of its cancer candidates.The Nabla-Takeda deal value could reach more than $1 billion biobucks if future milestones are met. Nabla’s Joint Atomic Model will work across Takeda’s early-stage programs to generate new antibodies, multispecifics and other therapeutic targets simultaneously. Takeda’s Innovent partnership could result in $10.2 billion in milestones and is meant to bridge the steep patent cliff created by the company’s inflammatory bowel disease drug Entyvio. The deal gives Takeda the rights to two Innovent candidates that have shown promise in fighting tumors in clinical studies.

Takeda has signed another AI-focused deal, this time focusing on small molecules with the San Diego biotech Iambic. The two announced the multiyear deal Monday, working on undisclosed targets in the oncology, gastrointestinal and inflammation areas. Takeda could pay up to $1.7 billion in potential milestones, and an Iambic spokesperson tells Endpoints News it includes upfront and research cost payments in the double-digit millions. Further financial details weren’t disclosed. In an interview, Iambic CEO Tom Miller said his startup’s focus on building its pipeline, instead of just AI models, led to “strong alignment” with Takeda. Iambic has delivered promising yet early, initial human data with its lead drug, the HER2-targeting molecule IAM1363, which is being tested in cancer. The 120-employee company is on track to start human testing this year with two more molecules as well. That focus on moving drug programs ahead has been central to Iambic’s pitch in the crowded, often noisy world of AI-focused biotech. “They see we can actually create drugs like IAM1363,” Miller said. “It’s not an AI pitch deck with a lot of hopes and dreams. It’s more about allowing the molecules that we create to speak on behalf of the platform.” Takeda’s chief scientific officer Chris Arendt said in a statement that Iambic’s platform could help “de-risk candidate selection, improve probability of success, and more quickly advance select programs from early project start to IND.” Beyond small molecules, Takeda expanded its partnership with Cambridge, MA-based Nabla Bio last fall for using its AI platform to design protein-based therapeutics. Iambic has ongoing research collaborations with Lundbeck and Revolution Medicines . Miller said the Takeda deal takes elements of both deals, which respectively focused on discovery and customizing AI models with data for specific research areas. Miller said Iambic has drawn additional pharma interest in partnerships out of last month’s JP Morgan Healthcare Conference, but will be selective in what it decides to do. “We’re not a do-a-lot-of-partnerships company,” Miller said. “That bar gets harder with every deal you do, because it has to offer differentiated value versus what you’re already doing with existing partnerships.” Some of Iambic’s key models include NeuralPLexer for predicting protein-drug structures and interactions, and Enchant for making a range of clinical predictions based on a molecule . The company was co-founded in 2019 by two academic chemists, Miller and CTO Fred Manby . For the year ahead, Miller said the company has runway through 2028 after recently raising $20 million from the Ireland Strategic Investment Fund, bringing its total raised to more than $320 million. He added that the biotech expects to enroll 100 to 150 more patients in its expanding HER2 clinical trials while bringing its next two drugs into the clinic.