Clinical trials are where most drugs go to die, but they don’t need to stay gathering dust on the shelves for an eternity or ultimately disintegrate.
Derisking is always important in drug development, especially when dealing with the central nervous system (CNS), which has experienced lower success rates historically and, because of its complexity, takes longer to develop than other drug classes. A review from the National Institute of Aging (NIA) reported that the success rate of neuropsychiatric drug candidates entering human testing and consequently being approved is dramatically lower (8.2%) than for all drugs combined (15%).
According to the Tufts Center for the Study of Drug Development, from 2000–2017, developing drugs to treat CNS disorders required, on average, 20% more development time than other drugs that won marketing approval in the United States and took 38% longer to win approval. One review stated that the time from molecule to drug on the market for CNS drugs is typically 12–15 years and that the cost of developing a new drug is typically $10–15 billion.
Spending billions of dollars over the decade can ultimately result in a drug that gets shelved because these compounds missed their endpoints for whatever reason. So, plenty of compounds are on the shelf that can be reexamined and redesigned for another indication, making the pathway to market a lot shorter and more cost-effective.
Friso Postma, PhD, VP of Artificial Intelligence Drug Discovery at BioXcel Therapeutics, a publicly traded biotech in New Haven, Connecticut, has dedicated the past three years or more to developing a comprehensive artificial intelligence (AI) platform for repurposing pharmaceuticals. Postma’s research team at BioXcel has been at the forefront of employing AI to re-engineer a pharmaceutical compound focusing on the nervous system.
“There’s probably a drug out there for most indications in neuroscience, but it’s not known,” said Postma. “We use AI to leverage these compounds and identify opportunities to create value.”
Augmented intelligence
The inefficiency of drug development in neurological and psychiatric conditions is no secret. Over the past 50 years, there hasn’t been much innovation in drugging neuropsychiatric disorders, and often the anxiolytics, anti-psychotics, and antidepressants that were introduced decades ago are either still in use or have been refined.
Postma believes this lack of innovation occurs for several reasons: the sheer amount of information, the complexity of neuropsychiatric circuits, symptoms, and behaviors, and the lack of objective endpoints. On top of that, there’s bad data—according to a news article from Nature, more than 10,000 research papers were retracted in 2023, breaking previous annual records.
“There are over two million publications/year in life sciences, which means two or three papers are published every minute—nobody can read that quickly!” said Postma. And there’s p-value hacking, falsified data, and all sorts of problems that plague modern science. There’s noise, and you have to be able to read through that. To make sense of all that information and detect the signal from the noise, we use AI.”
When Postma says BioXcel is using AI, he is not referring to artificial intelligence—rather, he prefers to call it augmented intelligence. That is because he insists it’s not a black box but a composite of AI tools that requires immense human supervision.
Postma led a team to develop an augmented intelligence platform called NovareAI for AI-based drug re-innovation by focusing efforts onto many of the brain’s highly conserved clusters of neurons located in the brainstem and cerebral hemispheres, also known as nuclei, which have well-defined conserved behaviors that have been modeled preclinically, associated with psychiatric indications, and tested in clinical trials. The reason for taking this approach is to make sure that there are established preclinical and clinical endpoints for candidate drugs before even embarking on the re-innovation endeavor to avoid making the mistake of investing millions, even billions of dollars, into a drug without a means of getting a proper readout necessary for FDA approval.
Ultra-speed readers
To sort through all the literature available to pull out information connecting compounds, neural circuits, behaviors, and indications, NovareAI uses “knowledge graphs” to mine the literature, organizing data from various sources and capturing information about entities that can connect compounds, neural circuits, behaviors, and indications. These AI systems organize data from various sources and capture information about entities, and create connections between them. Knowledge graphs include structured information that AI systems can use for various functions, including information retrieval, recommendation systems and question-answering, which is perfect for the types of queries Postma is interested in.
Postma said that some of the information that is fed into the knowledge graph is well structured, whereas there are lots of unstructured data, which requires a whole layer of tools like natural language processes and large language models to make sense of it all.
“You need to use words to connect the dots,” said Postma. “The ontologies and the harmonization of the terms and the entities are things that we pay a lot of attention to, and they are one of the aspects where a lot of creativity comes into play. You cannot just dump massive amounts of data and think you will get anything meaningful from it. As a matter of fact, the knowledge graph that we’re building has millions and millions of nodes, but it’s not huge compared to many of the other approaches that you see sometimes. It is, however, geared toward the questions we are asking. Our questions all concern drug re-innovation, which is a niche type of environment.”
That means they avoid the new chemical entity (NCE) domain entirely, which requires a slew of predictions to identify some molecules that may bind to a target. Instead, NovareAI is plugging away every hour of every day, trying to detect failed phase two and three assets it can quickly innovate upon. If they get a successful hit, they have just leapfrogged a handful of years in the drug development process and saved a whole bunch of cash.
Netflix for drug re-innovation
Postma and his team can ask questions from many different angles. They can approach the drug repurposing problem from the indication side, meaning picking an indication and then connecting it to a circuit, a target, and, ultimately, a compound.
Postma said, “It’s the same thing that Netflix does—based on existing information, can you predict what somebody might like regarding their next movie? In our case, we ask for recommendations for novel or existing links of compounds to an indication.”
Conversely, NovareAI can approach the drug re-innovation problem from the compound side, which Postma says can be done with “really good compounds”—ones that can be delivered to the brain quickly, are water-soluble, and can be reformulated. From there, it’s all about working the connections in the opposite direction to pin down a druggable target that can modulate an indication.
Take dexmedetomidine (Precedex), for example, which was originally formulated as an intravenously administered non-opioid drug to manage pain and sedation in the intensive care unit and operating room. Over time, the usage of dexmedetomidine expanded to off-label uses, including treatment and prevention of delirium and treatment of alcohol withdrawal. (Dexmedetomidine has also been used in peripheral nerve blocks to prolong the duration of analgesia.) Dexmedetomidine has qualities that are sympatholytic—it blocks nerves from the sympathetic nervous system (“fight or flight”), which can treat anxiety, such as generalized anxiety disorder, panic disorder, and post-traumatic stress disorder (PTSD).
Thanks to dexmedetomidine’s ability to act as a sympatholytic, BioXcel Therapeutics was able to find it as a possible match for treating adults with schizophrenia or bipolar I or II disorder who are highly agitated. BioXcel changed dexmedetomidine into IGALMI, a film that is placed under the tongue and used to treat extreme agitation. BioXcel changed dexmedetomidine into IGALMI, a film that is placed under the tongue and used to treat extreme agitation.
“With IGALMI, we went from first-in-human to approval in less than four years—that’s quite dramatically different from what you would see with a [new chemical entity],” said Postma. “You’d have to go through each selection, which takes 10–12 years and costs an average of $1.2 billion. There’s a tremendous upside here.”
BioXcel is also working on BXCL502, a new anti-stress drug that is not an antipsychotic that is being studied as a feasible way to treat chronic agitation in dementia, and BXCL503, which could potentially treat Alzheimer’s disease-related symptoms unrelated to agitation, such as apathy.
Jumping the line
Just because NovareAI gets a hit doesn’t mean that BioXcel can jump immediately into late-stage clinical trials, let alone submissions for approval—preclinical validation is still required. To save resources, BioXcel isn’t devoted to devising models from scratch but instead turns to existing models that hold translational value.
“If those models do not exist, then I can’t test my concept, and that will be deprioritized,” said Postma. “So, the ability to be able to do pre-clinical validation is one of the first filters that we set in prioritizing our list of concepts.”
But it doesn’t stop there. Postma said that another anticipatory filter is whether IP can be generated—if there’s no IP, there’s no upside for the company. For example, dexmedetomidine’s formulation, Precedex, was used intravenously, but IGALMI is a sublingual film for fast-acting responses because it bypasses first-order metabolism. Thanks to dexmedetomidine’s ability to act as a sympatholytic, the predecessor platform to NovareAI was able to find it as a possible match for treating adults with schizophrenia or bipolar I or II disorder who are highly agitated.
Other filters for selecting drug candidates for reformulation include the existence of clinically validated endpoints, bona fide regulatory pathways, competition, and market size.
According to Postma, NovareAI has evolved into something that holistically addresses drug re-innovation as a process from concept to market. It also illustrates that sometimes bigger is not better regarding company size.
“You have large companies out there that have a core unit of data engineers, a core unit of data scientists, and a core unit of AI engineers, and they need to interact, and at one point, it gets siloed—it gets too big, and nothing really happens,” said Postma. “What we do is very actionable and concrete. We have a small team, but we are nimble and can address all these facets of the drug renovation process. It is a constant process of educating each other about how you could maximize the output of the original AI.”
Striking the sweet spot with the balance of biological and artificial intelligence is at the core of BioXcel’s mission to develop medicines in neuroscience. Whether this approach can really be used to develop transformative, blockbuster drugs from the re-innovation of past failures remains to be seen.
Medicines sitting on shelves aren’t just going to grow legs, walk into a clinic, and start to save lives. It will take approaches like BioXcel’s to sort through the massive haystack of potential pharmaceutical needles that can thread gaps in medicine for untreated indications, and BioXcel’s approach could help to pull out a handful of them that relate to the nervous system.
