
Drug development is among the slowest, most failure-prone processes in modern science, with about 90% of drug candidates never reaching the market. Today, artificial intelligence methods have accelerated the first step — plucking promising molecules out of endless possibilities — but countless challenges remain. A successful drug must be not only safe and effective, but also able to bypass the body’s defenses and reach the right target.
Most drug candidates fail such optimizations. That’s where a new research institute at the University of Washington has focused its attention. Housed in the UW School of Pharmacy, the Institute for Innovations in Drug Delivery and Disposition (I2D3) brings together experts in artificial intelligence, drug discovery, pharmacology, data science and biotechnology to ease the bottleneck between promising molecules and successful drugs.
The Institute opened in July 2026 and is led by three UW faculty members: Gaurav Bhardwaj, an associate professor of medicinal chemistry who oversees the Institute’s AI-enabled molecular design; Nina Isoherranen, the Milo Gibaldi Chair of Pharmaceutics and expert in drug metabolism and disposition; and Marco Pravetoni, a professor of psychiatry and behavioral science in the UW School of Medicine, who leads drug discovery, translation and commercialization efforts.
UW News spoke with the three co-directors about why drug candidates fail, how AI is speeding drug development and how I2D3 hopes to help get drugs to market more quickly.
What separates a promising molecule from a full-fledged drug? What properties need to be considered, and how can a developer work toward them?
Gaurav Bhardwaj: It really depends on the disease indication you are targeting and the therapeutic modality. Let’s say you have a promising molecule that interacts with the disease-causing protein. Delivery becomes equally important — do we need an orally delivered drug? Do we need to cross the blood-brain barrier? If the disease requires daily dosing, then injectable or IV methods aren’t optimal. If it’s delivered orally, then the molecule needs to be able to get across the gut barrier, and also needs to be stable enough that it doesn’t get chewed up by the body. It also needs to stay in the body for a reasonable time. A successful drug molecule has to meet all these and more criteria, and ultimately all these criteria are encoded by the sequence and structure of the molecule.
The Institute is devoted to aspects of drug development that are often overlooked. What problem do you see the Institute being able to help solve?
GB: Traditional drug discovery and development is a trial-and-error-based process. Either you find a useful molecule in nature and spend years optimizing it for human use, or you create many random combinations of molecules and hope that one of them has the function you need. Both of these approaches are highly unsuccessful, which has created a bottleneck.
Now the field is also focusing on an idea called rational drug design. It started long before AI but is now becoming even more common. People are using AI methods to design new molecules. However, a lot of that work has focused on the first step — finding a molecule that binds to a specific protein, or has a specific function in the body. That’s still not a drug, it’s just more candidates.
The bottleneck has now shifted. It’s no longer finding that first molecule, but now, how do you add all the other drug-like properties? That’s what the Institute is trying to do. Let’s build the models that ultimately make molecules that are going to be successful all the way through the drug development pipeline.
Marco Pravetoni: I see our work also as accelerating discovery. I work on substance use disorders, and my lab develops vaccines, antibodies and next-generation antibody-like molecules that target drugs in the body. With these new tools, instead of working to design 10 antibody candidates in a lab, we could design 1,000 or more, and then we can accumulate enough data to reduce any risks, so that what we bring to clinical trials is more likely to be successful. AI can do a lot of that.
How can you make it more likely that a drug candidate succeeds in trials?
Nina Isoherranen: Part of it is predicting what’s going to happen to a drug in humans before it’s ever given to humans. That should increase the success rate and eliminate the waste of doing a lot of unsuccessful trials.
We can also build machine learning and AI approaches to predict drug disposition in an individual person. What we talk about today are ‘digital twins,’ which refers to a computational model of the individual patient and their characteristics. For example, how does your kidney function? What is your body mass index? And so forth. Then we generate a digital version of you. We can then predict how a certain drug would behave in your body and build the best strategy.
There’s also an access-to-treatment question here. Pregnancy is a great example — we often don’t know how drugs work in pregnant women because we’ve never done trials. To be safe, we say that pregnant people shouldn’t take those drugs, but that means they don’t have access to a potentially hugely beneficial medication. If we can use AI and machine learning to predict how pregnant people respond to medications and how their bodies handle drugs differently from nonpregnant people we can make more medications accessible
Now with AI and machine learning, I think we can get to a place where we can truly sample the full space of possibilities.
How can the methods you’re building help with these individualized treatments?
NI: We know that drugs behave differently in different people. Even if we give them the exact same drugs and concentrations, people may still have different responses because of factors inherent to our bodies.
During drug development the candidate drug needs to be studied to see responses in different populations. Before you get a drug approved, you need to understand how liver disease, for example, is going to change exposure to that drug and whether you need to change the dosing. There’s a lot of guidance on drug interactions. Pharmacists manage drug interactions all the time, but it gets very complicated when you combine multiple patient factors. Now, if we have good predictive tools, we can predict what’s going to happen without having to do trials.
The ultimate goal here is to be able to predict, using model computational tools, what’s going to happen in individual humans before you ever give them a drug. What’s the right dose? The right timing?
UW has established itself as a leader in these fields already. I’m thinking especially of the UW Medicine Institute for Protein Design, whose director, David Baker, recently won the Nobel Prize in Chemistry. How does I2D3 fit into the broader UW ecosystem?
MP: IPD is a world leader in designing novel proteins, and the UW also has outstanding capabilities in clinical testing and implementation through the Institute of Translational Health Sciences. However, there remains a critical translational space between discovery and clinical application — one that focuses on the pharmaceutical development needed to turn promising innovations into viable therapeutic products. That’s where I2D3 can play a leading role.
For example, when researchers at IPD develop a new protein, I2D3 can partner with them early to address formulation, manufacturability, stability, delivery, and other key pharmaceutical considerations that are essential for advancing a discovery toward the clinic and ultimately the marketplace. I2D3 would serve as a core translational partner, helping bridge the gap between innovation and implementation.
IPD brings unmatched strengths in protein design, ITHS provides expertise in clinical translation, and I2D3 contributes the drug development and pharmaceutical sciences capabilities needed to move discoveries across the translational continuum. Together, these organizations can create a powerful and highly integrated ecosystem.
For more information, visit i2d3.washington.edu. To reach the researchers, contact Alden Woods at acwoods@uw.edu.












