To develop new and better peptides, the short amino-acid strings behind medicines like GLP-1 drugs, researchers have used AI to generate candidates and to predict their properties .
However, merging these capabilities into a system that generates peptides likely to activate or block specific targets has proven difficult. In part, this is due to the vast number of possible peptides, but also because predicting how readily a peptide will bind to a target — like G protein-coupled receptors (GPCRs), a family of cell-surface proteins targeted by about one-third of approved drugs — is easier than simultaneously forecasting what effect that binding will have.
Now, researchers at the University of Pennsylvania and The Chinese University of Hong Kong have created TD3B, an AI framework that guides peptide generation toward candidates predicted to have a desired effect. The results, which focus on GPCRs, are described in a paper presented as a Spotlight at the 2026 International Conference on Machine Learning .
The advance could accelerate drug development by making it easier to design peptide drug candidates likely to have particular effects, such as improving GLP-1-related therapies for weight loss or diabetes, quieting brain signals involved in addiction or helping immune cells fight cancer, rather than generating new candidates first and testing them later to discover what they do.
"What matters is not only whether a molecule reaches the right target, but what it tells that target to do," says Pranam Chatterjee , Africk-Lesley Distinguished Scholar of Innovation in Engineering, Assistant Professor in Bioengineering (BE) and in Computer and Information Science (CIS), and senior author of the new study. "With TD3B, we are beginning to build that directionality into the design process itself."
Why Direction Matters
Cells constantly receive signals from their surroundings, such as hormones, neurotransmitters and other molecules that tell them how to behave. Embedded in cell membranes, GPCRs function like molecular doorbells. When the right molecule binds to one of these receptors, the receptor relays a message inside the cell.
This is why designing peptides that are likely to do more than just bind to a particular receptor matters. While some drugs act as "agonists," ringing the doorbell and spurring cellular activity, others serve as "antagonists," covering the doorbell so that no signals get through.
"It's critical to know whether a given peptide is expected to behave like an agonist or antagonist," says Aastha Pal , a doctoral student in Bioengineering and the study's co-first author. "A peptide might bind to the right receptor, but if it pushes that receptor in the wrong direction, the peptide could have the opposite of the desired effect."
That distinction is the reason GLP-1 drugs work. As agonists, they activate the GLP-1 receptor, found in the pancreas, gut and brain, triggering signals that help regulate appetite, weight and blood sugar. Drugs that bound to the same receptor but acted as antagonists, blocking rather than activating the signal, would not have the same effect.
Building TD3B
To incorporate directionality into AI peptide generation, the team had to solve three problems at once. TD3B — short for Transition-Directed Discrete Diffusion for allosteric Binder design — needed to generate plausible peptide candidates, predict whether those candidates were likely to bind to the desired receptor and then determine whether binding would activate or deactivate the associated cellular machinery.
The challenge is a little like searching a vast forest for a medicinal plant with a very specific effect. "The key was deciding what counted as a successful candidate," says Hanqun Cao , a doctoral student at The Chinese University of Hong Kong and the study's other co-first author. "Rather than rewarding peptides simply because they looked realistic or were predicted to bind, we only gave the strongest rewards to candidates that were predicted to both bind the target and produce the biological effect we wanted."
Technically speaking, TD3B combines three key subsystems. A machine-learning model known as the "Direction Oracle" predicts how a given peptide and receptor will interact. Then, a "gated reward" system gives the strongest scores only to candidates predicted to bind to the target and push it in the desired direction. Finally, a "training buffer" stores high-scoring candidates and uses them to guide future rounds of generation.
"A key idea behind the framework is that the model learns from its own best discoveries," says Sophia Tang , an Computer Science and Statistics undergraduate at the University of Pennsylvania, co-author of the study and lead developer of TR2-D2 , the core AI framework on which TD3B builds. "As it finds peptides that are more likely to produce the desired effect, TD3B uses those examples to guide the next round of generation, making each step more productive."
Testing TD3B
In addition to evaluating the Direction Oracle, which achieved 93% accuracy in distinguishing agonist- and antagonist-like interactions, the researchers compared TD3B-generated candidates with known interaction patterns for GLP-1 drugs, a process similar to comparing a newly made key to one known to open a particular door.
In computational structural analyses, TD3B's predicted agonists contacted key activation sites on the GLP-1 receptor, including those known to be essential for full agonist activity. The model's predicted antagonists, by contrast, avoided those same activation-related sites.
"That was one of the clearest signs that the model was learning something meaningful about direction," says Pal. "We did not tell TD3B to bind to those exact sites. We asked it for an agonist or an antagonist, and then we looked at the structures and saw that the agonist-like candidates were making interactions associated with activation, while the antagonist-like candidates were not."
The researchers saw a similar pattern with tests involving the orexin 1 receptor , or OX1R, a GPCR involved in sleep, wakefulness and reward-related behavior that may play a role in conditions as varied as insomnia and substance use disorder.
Future Directions
The researchers are currently synthesizing TD3B-generated peptide drug candidates and running tests with them in the lab, going beyond computational simulations to in vitro and in vivo experiments.
"We've already shown that TD3B can generate peptide drug candidates likely to bind to target sites and push them toward activation or shut them down," says Chatterjee. "The next step is to test whether those predictions hold up experimentally. If they do, this could open the door to designing peptide medicines based on the therapeutic effects we want to achieve, not just the proteins we want to bind."
This study was conducted at the University of Pennsylvania School of Engineering and Applied Science (Penn Engineering) and The Chinese University of Hong Kong (CUHK) and supported by the High-throughput Institute for Discovery (HIT-ID) at the University of Pennsylvania and the Research Grants Council of the Hong Kong Special Administrative Region, China, under Project T45-401/22-N.
Additional co-authors include Yinuo Zhang of Penn Engineering and Jingjie Zhang and Pheng Ann Heng of CUHK.
