OR WAIT null SECS
The application of artificial intelligence is steadily finding a home in the pharma industry, but these tools may have unprecedented potential in a more nascent niche along the life sciences continuum: drug development.
Over the past year, artificial intelligence (AI) has risen to the forefront of the tech industry. While the technology has existed for some time now, recent advancements have pushed it into the limelight. With the release of generative programs such as ChatGPT, workers across every industry began wondering how they could use AI-based programs to improve their work.
In the pharmaceutical and life sciences industries, the obvious answers revolved around data collection and management. This made the technology attractive to members of the sales and marketing teams, who found a variety of ways to put it to use. It’s even been implemented in programs to assist doctors in making diagnoses.
More recently, a less obvious, but highly effective, use for AI has presented itself.
Pharmaceutical Executive® spoke with Frank Yocca, PhD, chief scientific officer at BioXcel Therapeutics, about how his company developed a strategy to use AI during the drug development process. Yocca described BioXcel as a "re-innovation" organization, meaning that it takes drugs that are or have been in Phase II or Phase III clinical trials (or even some that have already made it to market) and finds new uses for them. In the case of BioXcel, it used AI to identify a new indication for Precedex (dexmedetomidine), which is used in surgical units as an anesthetic. Through its use of AI, the company advanced from intravenous first-in-human to FDA approval with a novel sublingual film to commercial launch in under four years. Last year, FDA approved BioXcel's first drug, Igalmi (a sublingual film formulation of dexmedetomidine), to treat acute agitation in adult patients with schizophrenia or bipolar disorders in medical settings.
Prior to developing its current strategy, one of the main challenges the company considered is the fact that the drug success rate in areas like neurosciences (where Yocca’s career work was focused) is low. For that reason, he says, many organizations moved away from neuroscience research.
In 2015, Yocca started working on a strategy to improve those odds of success.
“We came up with a unique strategy to use AI to re-innovate drugs,” he says. “We also had to determine what kind of molecules we would choose from AI to develop, how to do so, and [whether] we could do something to increase the probability of success. We decided to start out focusing on Phase II drugs. It can take six to seven years and cost up to $150 million to get through Phase II. Instead, we said, ‘Let’s find an area that we think has an unmet medical need and focus AI on it; and, hopefully, it can point us in a direction where we can innovate something and start out in Phase II.’”
Yocca explains that it’s important to realize that AI is not a black box that simply provides answers. It’s a tool, and the results it generates greatly depend on the type and quality of data being put into it relevant to the specific questions being asked. To get the best use of AI programs, Yocca learned that the first step was to identify the problems that he needed to address specific to the drug re-innovation process.
The result is a unique platform that Yocca refers to as containing "composite AI" because it can take multiple data sets into account while making an evaluation.
“Our AI platform helps in obtaining information about drugs, targets, and indications, which are all connected in labeled property [or knowledge] graphs,” he says. “This is the workplace for AI. The platform also includes machine and deep learning methodologies to learn the drug-like behavior of compounds.
Yocca notes that BioXcel's compendium includes all drugs that exist, so the platform helps narrow down the list of potential candidates. "We also use natural language processing to extract relevant information from text as well as graph-based data science, including methods to detect hidden information in the knowledge graphs," he says. "That’s where you find the gold; so, you must get into the knowledge graph and look at it from a three-dimensional space to see where unique connections between these targets and systems, particularly in neuroscience, exist.”
Working in neuroscience drug development presents unique challenges. For example, Yocca explains, unlike oncology, where the focus is on targeting and eradicating tumors, neuroscience deals with a broad range of psychiatric and neurologic diseases with multiple symptoms. This often means that patients have a multitude of different systems being impacted.
Typically, the most successful drugs in this area are what Yocca refers to as "dirty drugs." These medications have multiple mechanisms. Drugs that focus on more specific systems have less of a chance of being successful because they might not be normalizing the systems causing the disease and thus fail to address the symptom being targeted for treatment.
This is one reason the neuroscience setting works so well for re-innovation. Since drugs often impact multiple systems, there can be hidden uses for them that the original developers didn’t intend or realize.
“We’re working on recommendation systems to reveal hidden information in these knowledge graphs, using techniques such as matrix factorization,” says Yocca. “This is a similar technology that companies like Netflix use to predict what other movies you may like after having a sample of what you’ve been watching. Ultimately, we look to predict what drug can be re-innovated for which innovation. We use AI to augment that decision process when it comes down to drug innovation, along with big data (structured and unstructured). In essence, we address the problem of finding the signal amongst all the noise, using AI.”
Without AI, researchers like Yocca would have to process massive amounts of data on their own. Considering that more than two million papers are published each year in the life sciences, at a rate of about two per minute, that can quickly become overwhelming.
To get AI to use the data accurately, the user needs to appropriately contextualize it. For Yocca, that means carefully refining the questions being asked. While it’s true that an AI model is only as good as the information fed into it, it’s also true that the results are only as good as the questions presented to it. To that end, much of Yocca’s work is focused on determining the correct questions to pose.
“We’re also looking at ways to leverage and fine-tune large language models, such as GPT (generative pre-trained transformer), for internal purposes,” he says, “and we’ve implemented the bidirectional encoder representations from transformers (BERT) model, specifically, that’s trained on biomedical data. We’re deploying our composite AI platform across the value chain (discovery, development, regulatory, etc.) and all the aspects that you need to have information on to develop a new drug.”
One of the first questions in any discussion about AI is whether it is being used to replace workers. In the life sciences, and specifically in the drug development process, this question would refer to doctors and researchers.
Yocca explains that, to some extent, AI is being used to sort and quantify data, assigning relevance. Meanwhile, the final analysis is primarily still being conducted by humans. To be more specific, Yocca points out that AI is primarily focused on finding outlying pieces of data.
“When we’re looking for a new molecule, we’re looking for something that’s unique,” he says. “That’s why I talk about representing data in a three-dimensional space because you’re looking at circuits in neuroscience, and you don’t know how they come together. With the information that AI provides, you get connections in unexpected directions, and you start following it to see where AI is going. You then must decide whether it makes sense and if it’s possible to affect the physiology of this system by going in that direction. That’s when the neuroscientists and data scientists decide to set up a study to test it.”
AI is also being deployed in clinical situations. According to Yocca, it’s used to analyze data on drugs that have already been approved and focus on which patients do or don’t respond to the therapy. The hope is that, when enough data is collected, AI can start to identify patterns and researchers can predict which patients will respond to the drug before giving it to them.
Analyzing this data doesn’t just determine what drugs to give to patients, however. It’s also being used to optimize the best ways to administer medication.
“We’re looking at which hospitals have the most patients that come in with agitation related to schizophrenia or bipolar disorders,” says Yocca regarding BioXcel's focus. “There are also a lot of assaults that happen in hospitals. When a patient comes into a hospital setting and is demonstrating moderate agitation, the situation at the hospital may cause an escalation. They’re in a situation that may seem strange to the patient; people are firing questions at them and then somebody pulls out a syringe. The next thing you know, you’re dealing with a hostile patient. Based on this data, we’ve learned that it may make sense to put [the] patient on medication right away, along with doing a verbal de-escalation.”
Yocca stresses the results from AI really depend on the person using it. Since his team consists of both data scientists and neuroscientists, they are able to work together and learn from each other.
His team is considering tapping AI-powered language tools like ChatGPT—platforms that may make a lot of headlines, but not necessarily for the right reasons. For example, ChatGPT is known to "hallucinate," which is when it creates brand new information instead of referencing real data. Someone asking the program to write a paper on Greek literature may get a result that includes references and citations to books and writings that don’t exist. For a college student attempting to cheat, this is a problem. However, for the life sciences industry, issues like this can mean life or death.
Yocca explains that companies are likely to start fine-tuning their own GPT—ones where they can control the data that goes in, the parameters by which the programs sort the data, and the questions that users are able to ask of the program.
“I understand the fears that people have with this, but it’s all about what you do with it,” he says. “It is hard to imagine a runaway effect with composite AI since we tailor our questions and match this to relevant data sets. Thus, in essence, you put controls on it to do the things that you want to do. The benefit of AI is that it helps you think in different directions. It puts you in situations that may be uncomfortable, but you may be about to do something interesting.”
It’s easy to get caught up in the excitement over new technologies, which is why it’s important to determine their limitations and understand their capabilities. Too often, people can fall into the trap of expecting too much from technology and then trying to use it in situations where it’s unable to produce good results.
It’s just as easy for people to assume limits on new technology that don’t exist. This can also prevent it from being used to its full potential. As Yocca says, the usefulness of AI depends on the person using it. By understanding how the technology works and what its strengths and weaknesses are, Yocca and his team at BioXcel developed a unique use case for it, and, in turn, uncovered a new approach to re-innovating medicine.
Mike Hollan is the assistant managing editor for Pharmaceutical Executive® and Medical Device and Technology®.