Researchers at the University of Pennsylvania in Philadelphia and The Chinese University of Hong Kong have introduced TD3B, an artificial intelligence framework that directs the computational design of peptides toward specific biological outcomes. Led by senior author Pranam Chatterjee, with co-first authors Aastha Pal and Hanqun Cao, the team developed a system that integrates target binding and functional directionality into a single generative process. The framework will be presented as a Spotlight at the 2026 International Conference on Machine Learning, addressing a longstanding bottleneck in peptide drug discovery. Peptide therapeutics, including GLP-1 receptor agonists for diabetes and weight management, function by either activating or blocking G protein-coupled receptors, or GPCRs. Designing molecules that consistently produce a desired agonist or antagonist effect has proven difficult due to the vast chemical space and the complexity of predicting functional outcomes. TD3B, short for Transition-Directed Discrete Diffusion for allosteric Binder design, overcomes this by coordinating three machine learning subsystems. A Direction Oracle forecasts the interaction dynamics between a candidate peptide and its receptor. A gated reward mechanism penalizes molecules that bind without delivering the intended functional response, while a training buffer stores high-performing candidates to iteratively refine subsequent generation rounds. This closed-loop architecture, further developed with undergraduate researcher Sophia Tang, allows the model to learn exclusively from successful directional outcomes. Computational validations demonstrate the framework’s precision. In blind tests targeting the GLP-1 receptor, TD3B-generated agonists naturally converged on known activation interfaces without explicit site instructions, while predicted antagonists avoided those regions. The model exhibited analogous behavior with the orexin 1 receptor, which regulates sleep and reward pathways. The underlying Direction Oracle achieved 93 percent accuracy in distinguishing agonist-like interactions from antagonist-like ones, confirming that the system captures meaningful biophysical principles rather than superficial binding patterns. The advance shifts peptide design from a trial-and-error discovery model to a purpose-driven engineering discipline. By embedding functional directionality directly into the generative loop, the framework dramatically reduces the candidate attrition rate that typically delays therapeutic development. Researchers are now synthesizing the AI-generated sequences for in vitro and in vivo characterization to validate computational predictions against empirical data. If experimental results align with simulations, TD3B could establish a new standard for precision peptide therapeutics, enabling rapid design of targeted treatments for metabolic disorders, substance use conditions, and neurological diseases. The work underscores a broader industry transition toward AI systems that optimize for specific biological functions rather than generic molecular feasibility.