[Adobe Stock]
Friso Postma, vice president of AI for drug discovery at BioXcel Therapeutics, prefers the term “augmented intelligence” to “artificial intelligence.” His company uses AI tools to support human experts in drug repurposing efforts.
Postma, who holds a Ph.D. in signal transduction from the Netherlands Cancer Institute, transitioned to AI from wearable digital health devices. “I quickly found out that it’s very important to be able to contextualize AI,” he explains.
Now, Postma, trained as a biologist and electrophysiologist, that experience in contextualizing AI has been instrumental in developing and continuously refining an end-to-end platform for drug re-innovation, focusing on neurological rare and psychiatric disorders. This platform, which he helped create, is now central to his ongoing work. “Interestingly, I’m not a data scientist or an AI engineer, and I don’t write code. But I lead a large multidisciplinary team with AI engineers and PhDs, both in Europe (Prague) and India,” he says. “It’s a position where I can combine my interests in data analysis and technology with biology to address unmet medical needs in neuropsychiatric disorders.”
Why making AI actionable depends in part on the questions you ask
Postma emphasizes that in today’s AI buzzword climate, the real value lies in its application. “AI by itself doesn’t do anything,” he said. “It’s not really that difficult anymore to do AI. What’s really difficult is making AI actionable — making it concrete.” Toward that end, “you have to be very clever as to how you’re going to deploy it,” he said. “That depends on the kind of questions you are asking.”
Friso Postma, Ph.D.
Established in 2005, New Haven, Conn.-based BioXcel Therapeutics has obtained FDA approval for the sublingual Igalmi (dexmedetomidine). AI supported the commercialization of the therapy, which is protected by a patent covering the innovative sublingual film formulation. By reformulating the intravenous drug into a sublingual film, BioXcel Therapeutics was able to secure new intellectual property protection, extending the commercial life of the drug beyond the original composition of matter patents.
Training an AI platform to identify ‘very good’ compounds
To assign novel indications to an existing compound involves multiple aspects. “One is identifying the compounds and their properties,” Postma explained. “Do they cross the blood-brain barrier?” he asked. “There’s a predictive element there as well.” BioXcel takes a multifaceted AI approach to identifying promising compounds. “We trained a neural network that can assign properties to compounds and determine whether they’re suitable,” he said.
After identifying promising compounds, the next crucial step is pinpointing a target. These targets need to be expressed in the body, and then researchers can investigate if they’re linked to a specific disease. “It’s important to understand that in neuroscience, we’re dealing with collections of symptoms,” Postma said. “Unlike some diseases, it’s usually not just one gene causing a neuropsychiatric disorder. Symptoms are essentially derivatives of behavior, and behavior is governed by neurocircuitry. So neurocircuitry is essential to our thinking about this.”
How an idea helped spark an AI platform
BioXcel’s AI strategy began with a simple observation. The company’s chief scientific officer, Frank Yocca, Ph.D., recalled that dexmedetomidine, the active pharmaceutical ingredient in Igalmi, was “a very good compound,” Postma recalled. “And then I would ask the question, ‘Well, what does that mean?’ So we created a platform, a system that looks and tries to identify ‘very good compounds.’”
The approach led BioXcel to develop a multifaceted AI system known as EvolverAI, including a knowledge graph, a type of knowledge base that uses a graph-structured data model to represent and process data. EvolverAI aims to accelerate the study of the complex relationships between diseases, molecular targets and drugs to identify novel therapeutics with the aim of re-innovating existing drugs.
Building the brains behind BioXcel’s AI
BioXcel’s approach involves looking at evolutionarily conserved neurocircuits that have particular targets in them, and then trying to see how that might affect behavior and, consequently, symptoms. “We’ve created a Knowledge Graph where we put all these things together. Then you get into graph-based neural networks that can subsequently try to predict, based on existing information, what targets may modify behavior and what other types of symptoms and behaviors could be affected,” Postma explained. “This is very similar to recommendation systems used by organizations like Netflix, where based on existing information about what you’ve been watching, they can predict what else you might like.”
The knowledge graph is a crucial part of BioXcel’s system, but it’s not the only component. “AI is a composite. EvolverAI, our platform, is a collection of different types of functionality that specifically tries to address all the different steps and facets involved in drug reinnovation,” Postma said. “We really operate in this niche.”
Data strategy is another pillar in successful AI deployments. That involves identifying the type of data required, but also organizing it. “If we have irrelevant data in our knowledge graph, we won’t be able to get the answers we need,” Postma said.
The process of gathering and organizing the data also involves a human element. “It’s an integrated effort between the biologists, the MDs, CMC people, regulatory people and the actual AI engineers,” Postma said.
An AI repurposing niche
BioXcel Therapeutics has relatively little competition in its strategy to use AI to enhance drug repurposing. There are an array of inputs in this hunt. “We actively screen clinical trials and SEC statements to identify promising phase 2 assets,” explains Dr. Postma. The company also uses natural language processing tools to scour academic publications, conference abstracts and and PubMed articles. “We have paradigms where we define ontologies and particular words and items that we’re interested in,” Postma said. The NLP system then retrieves abstracts that are relevant. “Ultimately, you will have to go through a process of manual curation, but the more clever you are about defining what it is you’re after, the better the result,” Postma added.
The process isn’t necessarily straightforward given the significant numbers of retracted academic papers and ongoing problems with skewed data in some papers. “There’s a lot of noise in the system,” Postma said. “One of the challenges of using AI is actually being able to detect the signal in the noise. The systems are constantly trying to retrieve information that we think is important. We look at it and then we incorporate it into our knowledge graph.”
Future possibilities
BioXcel Therapeutics is tapping its AI-driven drug repurposing platform to advance a pipeline of candidates targeting various disorders. With the recently approved Igalmi for acute agitation in schizophrenia and bipolar disorders, the company is exploring additional indications like Alzheimer’s disease agitation (BXCL501 in phase 3), opioid use disorder and PTSD (both in phase 2). Earlier-stage programs include BXCL502 for chronic agitation in dementia and candidates for apathy (BXCL503) and aggression (BXCL504) in dementia.
Dexmedetomidine may have potential in treating post-traumatic stress disorder (PTSD), a condition with limited treatment options. “Dexmedetomidine addresses a really distinct type of neurotransmitter system, which is norepinephrine,” Postma said. “We know it plays a role in stress, agitation and insomnia. You see that reflected in things like opioid use disorder, PTSD, and traumatic brain injury.”
By targeting the norepinephrine system, dexmedetomidine has the potential to reduce sympathetic hyperarousal and improve sleep quality. “So if you’re able to reduce the sympathetic hyperarousal and improve sleep, for instance, you will already be a long way towards these people feeling a little bit better,” Postma concluded.