Penn and Hong Kong Researchers Create AI Framework TD3B for Peptide Drug Design

Researchers at the University of Pennsylvania and The Chinese University of Hong Kong have developed TD3B, an artificial intelligence framework that generates peptide drug candidates and predicts their biological effects on cellular receptors. The model was presented as a Spotlight at the 2026…

Scientists at the University of Pennsylvania and The Chinese University of Hong Kong have unveiled a new artificial intelligence system named TD3B that can generate peptide sequences and forecast how those sequences will influence cellular receptors. This work was featured as a Spotlight presentation at the 2026 International Conference on Machine Learning, marking a notable step in the field of computational drug discovery. The core challenge the researchers tackled is designing molecules that do more than simply attach to a target; they want peptides that actively direct what the cell does next.

The Challenge of Functional Control in Peptide Design

Peptides are short chains of amino acids that already form the backbone of several existing medicines. The GLP-1 analogs used for diabetes and weight loss are one prominent example. Traditional artificial intelligence models in this space have largely focused on two separate tasks: generating peptide sequences on one hand, and predicting how strongly those sequences will bind to targets on the other. Many of those targets are G protein-coupled receptors, which mediate roughly one third of all drug actions. But binding alone does not tell the full story. The functional outcome is what matters most for treatment. A peptide can either activate a receptor, making it an agonist, or block it, making it an antagonist. Knowing which of these effects will happen is critical for therapeutic success.

How TD3B Works: Three Integrated Subsystems

TD3B brings together three core subsystems to handle this complexity. The first is called the Direction Oracle. It is a machine learning model that predicts how the interaction between a peptide and its receptor will translate into either activation or inhibition of that receptor. This gives the system a way to evaluate not just whether a peptide binds, but what it does after binding. The second component is a gated reward mechanism. This mechanism biases the generation process toward peptides that are predicted to both bind effectively and produce the desired functional effect. It acts as a sophisticated filter, going far beyond simple binding affinity scores. The third subsystem is a training buffer. This buffer takes the top performing peptide candidates and uses them to iteratively refine subsequent rounds of design. The result is a generative process that becomes progressively more targeted with each cycle.

Validation Across Receptor Families

The predictive capability of TD3B was put to the test through computational structural analyses involving the GLP-1 receptor. Agonist peptides that TD3B generated consistently engaged the parts of the receptor that are essential for activation. Antagonist peptides generated by the same model avoided those sites. The model was never explicitly told to target those locations. It learned to do so on its own. Parallel tests were run on the orexin 1 receptor, which is implicated in sleep regulation and addiction behaviors. Those tests showed similarly promising patterns, indicating that TD3B may work across different families of G protein-coupled receptors.

Implications and Next Steps

This method changes the early phases of peptide drug discovery by building directionality into the design process from the start. Instead of generating a large number of molecules and then screening them one by one for their effects, TD3B proactively focuses on candidates that are likely to have the desired therapeutic action. This precision could shorten the path from a computational design to a clinical candidate. The researchers believe it could open doors to more effective treatments for complex conditions such as diabetes, addiction, and cancer. The team is now synthesizing TD3B designed peptides for laboratory experiments. If those experiments confirm the AI's predictions, the framework could change how peptide medicines are conceived, moving beyond simple target engagement toward prescriptive modulation of cellular signaling pathways.

Pranam Chatterjee, the senior author of the study, commented on the significance of the advance. He said: "Designing molecules that not only find the right target but also control its behavior is the next frontier. TD3B marks a pivotal step in embedding this directionality into computational drug design." The research was supported by the High throughput Institute for Discovery at Penn and the Hong Kong Research Grants Council. The work highlights the synergy of artificial intelligence, structural biology, and medicinal chemistry in crafting next generation therapeutics.

The subject of the research is cells. The full article is titled "TD3B: Transition Directed Discrete Diffusion for Allosteric Binder Generation" and was published on July 6, 2026. The reference can be found at the OpenReview forum link provided by the team. Image credits go to Sylvia Zhang at Penn Engineering.

Peptides referenced: GLP-1.