Nikolay Dokholyan, a researcher at the University of Virginia, has been working on a problem that has silently defeated pharmaceutical science for decades. Finding molecules that appear promising on paper is the issue. There are many of those produced by the industry. The issue is that those molecules frequently don’t fit when they eventually come into contact with the proteins they are meant to target inside a real human body because the proteins have changed shape, much like a lock does when it is pressed against. The molecule was intended to be a statue. When it got there, something was in motion. Another billion dollars vanishes into the void between biological reality and laboratory promise as the drug fails and the trial collapses.
YuelDesign, YuelPocket, and YuelBond are a set of AI tools developed by Dokholyan in response to this, and their differences are truly significant. The majority of AI drug design systems currently in use treat proteins as frozen structures; they optimize a molecule to fit a shape that is never truly static in the body. Diffusion modeling, a type of artificial intelligence, is used by YuelDesign to design drug molecules while taking into consideration how proteins flex and shift during binding.
Key Facts & Context
| The Core Problem | Average cost of developing a new drug estimated at $2.6 billion or more; nearly 90% of new drugs fail when they reach human clinical trials — largely due to poor molecular binding predictions |
|---|---|
| UVA Breakthrough | University of Virginia researchers developed YuelDesign, YuelPocket & YuelBond — a suite of AI tools using diffusion models to design drug molecules while treating proteins as flexible, dynamic structures rather than frozen shapes |
| Why Flexibility Matters | Proteins change shape when a drug binds to them — a phenomenon called “induced fit.” Most existing AI drug tools ignore this, designing molecules for a static protein. YuelDesign accounts for real-time protein movement, producing far more accurate drug candidates |
| UCSF Protein Engineering | Professor Tanja Kortemme’s lab at UC San Francisco created the world’s first shape-shifting synthetic proteins using AI — described as “ChatGPT for proteins.” The work draws on NIH-funded data from the open-access Protein Data Bank, built over decades by global researchers |
| DeepMind AlphaFold 3 | DeepMind’s AlphaFold 3 solved a 50-year-old biological challenge by accurately predicting 3D protein structures — enabling researchers to design novel proteins for vaccines, cancer treatments, and pollution-eating enzymes |
| Nobel Prize Recognition | The 2024 Nobel Prize in Chemistry was awarded for using AI to design proteins — marking the first time AI-driven biological research received science’s highest honor |
| Industry Investment | Eli Lilly partnered with Encilico Medicine on AI-developed oral drugs in a deal potentially worth $2.75 billion; multiple pharma companies accelerating AI-assisted Phase I and II clinical candidates |
| AI Detection Capabilities | AI systems now detecting diseases from sleep patterns with up to 89% accuracy; predicting Alzheimer’s disease up to 10 years before clinical symptoms appear |
| Key Remaining Challenge | No purely AI-designed drug approved by the FDA yet; challenges include data-sharing gaps, intellectual property protections for algorithms, and the need to integrate biological lab experiments with computational outputs |
According to one of the project’s researchers, most tools create a key for a lock that is motionless, but theirs creates the key while the lock is in motion. It’s a more realistic metaphor than the majority of science communications. YuelDesign was the only tool that was able to capture the crucial structural alterations that take place during drug binding when the team tested it on CDK2, a well-known cancer-related protein. It was completely missed by the others.
This type of research is surrounded by a larger context of growing momentum and significant stakes. The average cost of developing a new medication is currently estimated to be $2.6 billion, and almost 90% of candidates that go through human trials still fail. Despite significant funding and sincere scientific efforts, those figures have hardly changed in decades. The pharmaceutical industry is full of people working against a problem that has consistently proven harder than it looks, not people who have given up. AI won’t be the answer. However, there is mounting evidence that it can significantly advance some of the most difficult aspects, especially the initial phases of finding promising molecular candidates and forecasting their behavior in biological settings, where both computational brute force and human intuition frequently fail.
Professor Tanja Kortemme’s lab at UCSF has been tackling the same fundamental problem in a different way by creating completely new proteins from scratch rather than refining already-existing drug molecules. Using AI models trained on data from the open-access Protein Data Bank, a repository developed over decades with funding from the National Science Foundation and the National Institutes of Health, her team recently produced what they claim to be the first shape-shifting synthetic proteins in history. She explains the potential in a way that is worth pondering: if you create proteins from scratch, you are no longer restricted to those found in nature. You can create something that targets a disease mechanism in a way that nothing currently on the market can, with characteristics that no natural protein possesses. She claims that the possibilities are virtually endless. This is the kind of assertion that typically elicits skepticism, but in this instance, it feels more like a thoughtful scientific evaluation than promotional language.
It appeared that the Nobel committee came to a similar conclusion. For the first time, AI-driven biological research was recognized with the 2024 Nobel Prize in Chemistry for using AI to design proteins. By correctly predicting the three-dimensional shape of proteins, DeepMind’s AlphaFold 3 has already resolved a 50-year-old structural biology problem, opening doors that scientists had been fighting for generations. Enzymes that degrade industrial pollutants, cancer treatments, and vaccines are already being developed using designer proteins. Declaring revolutions before the clinical evidence catches up has caused caution in the scientific community. However, the current rate of development is truly exceptional.
The pharmaceutical industry has responded by making investments, just like other industries do. Eli Lilly and Encilico Medicine reached an agreement to create AI-designed oral medications that could be valued at $2.75 billion. AI platforms are being used by several companies to expedite Phase I and II clinical candidates in ways that would have taken much longer using traditional methods. Although the exact date of the FDA’s approval of the first drug created entirely by AI is still unknown—that milestone hasn’t been reached yet—the companies making significant financial bets on it don’t seem to believe it will be far off.
As all of this progresses, it seems as though pharmaceutical research is on the verge of a real turning point, one that has been expected long enough that some skepticism is totally warranted but whose underlying science now appears more plausible than it has ever been. For a long time, the drug development pipeline has been flawed; it is costly, slow, and unreliable in ways that have serious repercussions for patients who are waiting for treatments that never quite materialize. AI won’t immediately resolve that. However, the problem of fitting the key to the moving lock is real, and researchers in labs in Charlottesville, San Francisco, and other places are getting closer to a solution than they have ever been.
