When Kristy Carpenter, a DBDS Ph.D. student, recently won the “Drug Repurposing and Repositioning Insights for Treating SUDs” Challenge sponsored by the National Institute on Drug Abuse (NIDA), it was the result of research that had unfolding for years, beginning at a surprising origin point. Her proposal, “Repurposing Telmisartan for Opioid Use Disorder” involves using a well-established hypertension medication to curb opioid usage, which, if successful, could help people struggling with opioid use disorder manage their condition.
As unlikely as it sounds, Carpenter’s work on this topic began with social media. Carpenter uses social media in a much different way than most of the world does. When she logs in to Reddit or Twitter, she’s not looking to catch up with friends, family or news — she’s looking for chatter about opioids, overdoses, and possible crises in a particular geographic area.
By using large language models, Carpenter employs them to search out certain patterns in usage, for example when an “H” signifies heroin depending on the context in which it’s written. By searching out frequency, location and a lexicon of other related terms, Carpenter can link “chatter” about opioids with overdose death rates within a community.
Coming straight to Stanford from MIT where she earned her undergraduate degree, Carpenter, who also secured a NRSA Individual Predoctoral Fellowship (F31) in 2024, began her time in DBDS with the idea of using AI for some form of drug discovery.
“I started out wanting to work on molecular structures, so a lot of my lab rotations in my first year were working with protein structure and RNA structure,” she explained. “I was always very interested in using AI for drug development and drug discovery. Then the interest in social media text only started in my second year when I wanted to explore biomedical data science more broadly.”
In her work with molecular structure, Carpenter began to focus on binding pockets on proteins in the human proteome, with the goal of finding where drugs could bind and potentially be repurposed. But in a different project, she was also looking at population-level social media text data in pursuit of aiding opioid addiction research. After years of working on these two projects in tandem, an opportunity to connect these different areas of research arose.
“What was really exciting about the NIDA drug repurposing challenge was it let me harmonize these two areas — because I didn’t start my PhD thinking that I would work on these two very different things at the same time,” she added.
That perspective allowed Carpenter to make use of her work with drug-binding pockets and consult with experts in addiction medicine that she connected with through her social media work to advise her on what a good therapy for opioid use disorder would look like.
“Then I was really able to blend these two halves, so it was it was super fun, very rewarding,” Carpenter said. “It really suited these two separate parts of my grad school life and brought them together.”
The social media data work was actively being pursued in Russ Altman’s lab when Carpenter joined several years ago. The research was spearheaded by then-Ph.D. student Adam Lavertu, and the response to it was so positive that a new grant was written and subsequently funded by NIDA to use social media text to try to identify trends in the opioid epidemic.
“The original plan was trying to forecast where overdoses might occur — not specific overdoses, but trends in overdoses like which areas of the country or areas of a given state might have some warning signs about coming overdoses. I want to emphasize that this is all at the population level; we’re not trying to surveil individuals,” Carpenter said. “I went to a NIDA-sponsored conference in Bethesda a little over a year ago and I presented a poster on this work; I just found it so compelling to talk to people who are working on opioid use, opioid addiction, overdoses trying to lessen the burden of this huge crisis in the country and in the world.”
It was that experience that caused Carpenter to decide to include more of an element on pain medicines and therapeutics for opioid use disorder in her Ph.D., which was then solely focused on the drug-binding pockets research.
“Something that people find surprising is that we don’t have to identify the context in which people are talking about opioids, so it doesn’t have to be the case that someone’s talking about their own drug use,” Carpenter explains. “They could be talking about drug use that they see in their community, they could be talking about policies on opioids that are happening in their community, they could be talking about drugs and pop culture or making jokes about drugs. There is a correlation between any mention of opioids and overdose death rates. So, it just seems that the more that people are talking about it, that indicates a greater presence of these drugs in the community.”
The unique problem about social media text is unlike when looking at the scientific literature, misspellings, slang terms and non-standard terminology are generally involved. So instead of only looking for exact matches of standard opioid-related keywords, Carpenter created a lexicon of slang and misspellings to use for searching social media text. Because LLMs such as GPT-4 have been trained on high volumes of internet data, they have seen a plethora of examples of how people talk on the internet. Carpenter used an LLM to generate a lot of proposed misspellings and slang terms, then used automated and manual filters to get rid of any erroneous terms. This process is faster and more scalable than asking an expert to come up with such a lexicon from scratch.
Carpenter is also currently working on using these GPT models to disambiguate between terms; for example, one slang term for heroin is just the letter H, although many of the uses of “H” online will not be in reference to heroin. But she sees a fix for that.
“Something that I’ve seen is that with these current large language models, they’re very good at determining when the letter H means heroin versus what it means any other number of things, “she said. “This next project of mine will hopefully be showing people that they can include these more ambiguous terms and get more signal from these varying noisy platforms.”
Her NIDA drug repurposing project, still in its early stages, began while Carpenter was looking at the binding pockets to which drugs bind to in the body.
“There is one protein in your brain called the mu opioid receptor (MOR) which is where these opioids such as fentanyl, codeine, and heroin primarily bind to,” Carpenter explains. “I was looking at the mu opioid receptors and my hypothesis was if I can find some protein that has a binding site that is highly similar to the mu opioid receptor binding site, the same drugs will probably bind these two binding sites.”
It’s a paradigm in drug repurposing, and Carpenter uses this as an example: If you have two locks that are very similar then probably the same key can fit into both of them.
Carpenter began looking for proteins that were known to bind drugs that already exist and are in use. It’s a way to save time in the drug development process.
“You find a drug that’s already shown to be safe, and that is already off patent— off patent is easier because then it’s much cheaper to produce, it’s already safe, it’s already in use, so that it can be used more quickly than developing all new drugs,” she said. “That’s the point of the drug repurposing.”
Carpenter looked at proteins that are currently known to be bound by existing drugs and identified those with highly similar binding pockets to the MOR binding pocket. She compiled all drugs that bind these pockets. This resulted in a lengthy list. A collaborator suggested that a successful medication to treat opioid use disorder would need to have a long half-life, meaning that it stays in the body for an extended period. Another consideration when choosing the right drug for repurposing was if it could cross the blood brain barrier, which is necessary for activity in the brain. After narrowing her list of potential repurposing candidates for these two criteria, Carpenter utilized molecular docking to see how well each drug candidate fit in the MOR binding pocket. Telmisartan, a medication used for hypertension, was one of the top two hits that performed much better than the other drugs that she screened.
“One thing that was really exciting is that when I looked at Telmisartan in the literature, there were already some studies on it that showed that it can have an effect in the brain,” Carpenter said. “There are previous studies in mice that show it protects against neurodegeneration — for example, Alzheimer’s — and there’s also one study in rodents that shows if you have rats that are addicted to alcohol and you give them telmisartan, it helps with their addiction to alcohol or their tendency to prefer alcohol. That made me really excited because that seems that this is real, it can have some sort of effect in the brain.”
The next phase for Carpenter is to collect this research into a paper.
“I would love to continue with this because if this is real, and not some spurious result, then it could hopefully be helpful or drive more research in this area,” she added.
As a fifth-year Ph.D. student. Carpenter’s time now is focused on completing her doctorate. She’s open to where her post-Stanford path takes her, whether it be industry — working at a pharma or biotech company — or academia. But the opportunity for computational tools, machine learning and AI to accelerate the drug discovery process is very appealing to her.
“It can take 10 years to develop a new drug, it costs a whole lot of money and there are many opportunities for computation to speed this up,” she said. “One very possible future is that I end up working at a biotech or pharma company. It’s a future I definitely see for myself. I found the opioid use disorder space and the addiction space very compelling and that’s not something that I expected coming into my PhD. It feels very satisfying and fulfilling to be working on a problem that is such a huge, huge issue today; probably one of the biggest issues in medicine in our country. I would love to continue working in this in this sphere, but we’ll see.”
—Laurie Notaro, DBDS Director of Communications