New AI Framework "TD3B" Designs Peptides With a Specific Biological Job in Mind

For years, AI-assisted peptide design has focused on two separate questions: which short amino acid chains are plausible drug candidates, and which targets they're likely to bind. A harder problem has remained largely unsolved — predicting not just whether a peptide binds a target, but what it tells that target to do once it gets there.

Researchers at the University of Pennsylvania and the Chinese University of Hong Kong say they've made real progress on that problem. Their new framework, called TD3B (Transition-Directed Discrete Diffusion for allosteric Binder design), generates peptide candidates aimed at a specific biological effect — pushing a receptor toward activation (agonism) or suppression (antagonism) — rather than simply optimizing for binding likelihood. The work was presented as a Spotlight paper at the 2026 International Conference on Machine Learning.

The study centers on G protein-coupled receptors (GPCRs), a large family of cell-surface proteins that mediate roughly a third of all approved drugs. GPCRs act like molecular doorbells: a bound molecule can either ring the bell (agonist) or hold it silent (antagonist), and the same receptor can produce opposite biological outcomes depending on which role a candidate peptide plays. Historically, generative models could suggest peptides likely to bind a target, but not reliably predict which of these two directions the interaction would push. For research teams screening peptide libraries, that gap has meant more downstream synthesis and testing to sort agonist-like candidates from antagonist-like ones.

According to the researchers, TD3B combines three subsystems: a "Direction Oracle" that predicts how a peptide-receptor pair will interact, a gated reward system that scores candidates only when they are predicted to both bind and produce the desired directional effect, and a training buffer that feeds high-scoring candidates back into future generation rounds. In testing, the Direction Oracle reached 93% accuracy distinguishing agonist-like from antagonist-like interactions. In computational structural analysis of GLP-1 receptor candidates, TD3B's predicted agonists contacted known activation sites, while its predicted antagonists avoided them — without ever being explicitly told which structural sites to target. The team observed a similar pattern testing candidates against OX1R, a receptor tied to sleep and reward-related signaling.

The researchers are now synthesizing TD3B-generated candidates for in vitro and in vivo lab testing to see whether the computational predictions hold up experimentally.

For research peptide suppliers and laboratories, tools like TD3B point toward a future where computational pre-screening narrows candidate pools before synthesis, potentially sharpening how research-grade peptide libraries are designed and prioritized for study.

For research use only. Not for human consumption.

What is TD3B?

TD3B is an AI framework developed by researchers at the University of Pennsylvania and Chinese University of Hong Kong that generates peptide candidates predicted to produce a specific effect — activating or blocking a target receptor — rather than only predicting binding likelihood.

What receptors did the researchers study?

The framework was tested primarily on GPCRs, including the GLP-1 receptor and the orexin 1 receptor (OX1R), both frequently studied targets in peptide and small-molecule drug research.

Has TD3B been tested in living systems?

Not yet. As of this study, validation has been computational and structural. The research team is currently synthesizing candidates for in vitro and in vivo laboratory testing.

Where was this research presented?

As a Spotlight paper at the 2026 International Conference on Machine Learning (ICML).

Medical Xpress: https://medicalxpress.com/news/2026-07-ai-tool-peptides.html

OpenReview paper (TD3B): https://openreview.net/forum?id=gPufROlvJF