Boston-based DeepCure may be an AI-focused biotech, but it sets itself apart by focusing its AI-powered platform on historically challenging drug targets that have eluded traditional approaches. “While AI in drug discovery is often viewed as a means to accelerate and reduce costs, our focus at DeepCure is on achieving true novelty in drug development to address targets that have eluded researchers for decades,” said CEO Kfir Schreiber.
The company’s first AI-generated drug candidate, DC-9476, a selective Brd4 BD2 inhibitor, embodies this approach, showing promise in preclinical models of autoimmune diseases like rheumatoid arthritis (RA) and Still’s disease. In RA models, DC-9476 demonstrated superior efficacy compared to standard treatments, including TNF-alpha inhibitors, IL-6 inhibitors, and the JAK inhibitor tofacitinib. DeepCure recently announced a collaboration with the Leeds Institute of Rheumatic and Musculoskeletal Medicine (LIRMM) to test DC-9476 in RA.
DeepCure’s AI capabilities extend beyond drug discovery to include innovations in chemical synthesis. Earlier this year, the company highlighted teh capabilities of its Inspired Chemistry platform, demonstrating its ability to synthesize nirmatrelvir, the active ingredient in Paxlovid, along with 56 analogs in a single, automated workflow.
Exploring new chemical spaces
A central pillar of DeepCure’s AI-based discovery approach is reinforcement learning (RL), an AI technique more often found in robotics and gaming applications. In DeepCure’s case, RL helps overcome inherent biases in traditional drug discovery. “At DeepCure, we use reinforcement learning alongside physics-based methods and other computational tools to tackle some of the most challenging unsolved problems in small molecule drug discovery,” said CEO Kfir Schreiber.
Kfir Schreiber
The challenge of exploring uncharted chemical spaces is central to DeepCure’s mission. With an estimated drug-like chemical space containing roughly 1060 molecules, medicinal chemists have only scratched the surface of exploring the universe of druggable compounds. “When you look at molecules, it’s not only that we have synthesized and tested a tiny fraction of the possible chemical space — it’s also that the molecules we have made are very clustered in specific areas of chemical space,” Schreiber said. This reality motivates DeepCure’s work to find novel compounds that have evaded traditional drug discovery methods.
In the hunt, DeepCure aims to eliminate what it terms “design biases,” Schreiber said. “We designed a generative AI tool that builds libraries on the fly for each specific target. This algorithm taps reinforcement learning, physics-based simulations and a few other tools to really eliminate design biases and enable true novelty.”
DC-9476: A case study in AI-driven discoverDeepCure
Brd4, a protein historically associated with dose-limiting toxicities, such as thrombocytopenia, has presented a significant challenge for drug developers. Previous attempts to target this protein via non-selective BET inhibitors have thus struggled with safety issues. To counter that problem, DeepCure focused on optimizing selectivity with their third-generation BRD4 (BD2) inhibitor, DC-9476. “We used our platform to identify a new way to interact with Brd4, a novel binding mode to achieve this selectivity,” explained Schreiber. “This allowed us to design new compounds that leverage this binding mode to be really selective.”
Promising preclinical results
DC-9476 has demonstrated promising results in preclinical studies. “We’ve shown that DC-9476 outperforms TNF-alpha inhibitors, IL-6 inhibitors, and JAK inhibitors in rodent RA models,” Schreiber noted. This superior efficacy in preclinical models of rheumatoid arthritis suggests the potential of DC-9476 as a novel treatment for autoimmune conditions.
The success of DC-9476 in preclinical studies is attributed, in part, to its unique mechanism of action. “By targeting this sort of ‘inflammatory checkpoint’ that stands at the center of many different inflammatory pathways, we have this very unique mechanism of action,” explained Schreiber. “It’s active in both T cells and macrophages, so both the adaptive and innate immune systems.” This dual action on both the adaptive and innate immune systems may offer a new therapeutic avenue for patients who haven’t responded well to existing treatments.
Future directions
Beyond its initial success with DC-9476, DeepCure’s AI-driven platform is poised to impact a range of therapeutic areas. The company’s recent collaboration with the Leeds Institute of Rheumatic and Musculoskeletal Medicine (LIRMM) exemplifies this potential. Together, they will evaluate DC-9476 in patients with rheumatoid arthritis (RA), particularly those who haven’t responded to existing therapies. The research will involve analyzing blood samples and joint biopsies from various patient subgroups, offering a deeper understanding of the drug’s efficacy and mechanism of action in a real-world setting.
This focus on addressing unmet medical needs is a cornerstone of DeepCure’s strategy. “If you have a tool like that,” explained Schreiber, referring to its AI platform, “the best way to apply it, in our opinion, is on targets where there is a very significant unmet need, but also a very big medicinal chemistry challenge that needs to be solved.” Rheumatoid arthritis, with over a million patients worldwide who don’t respond adequately to current treatments, represents a prime example of such a challenge.
While DC-9476 is DeepCure’s first AI-generated drug candidate to reach this stage of development, Schreiber emphasized that it’s “the first of many novel development candidates expected to emerge from our platform.” Outside of targeting BRD4-BD2, the company’s pipeline is also targeting STAT6 for asthma and atopic dermatitis as well as an undisclosed target for inflammatory-mediated indications.
“We’re excited about the road ahead and deeply committed to realizing the potential of AI for patients in need of better medicines,” Schreiber concluded.
Filed Under: Data science, Immunology, machine learning and AI
