TD3B: Transition-Directed Discrete Diffusion for Allosteric Binder Generation

Protein function is often controlled by ligands that bias the direction of state transitions, such as agonists and antagonists, rather than stabilizing a single conformation. This is especially important for clinically relevant G protein-coupled receptors (GPCRs), where therapeutic efficacy depends on functional directionality. Structure-based design methods optimize binding to static conformations and cannot represent non-reversible, directional effects or systematically distinguish agonist from antagonist behavior. To address this gap, we introduce Transition-Directed Discrete Diffusion for allostericBinder design (TD3B), a sequence-based generative framework that designs binders with specified agonist or antagonist behavior via a directional transition control objective. TD3B combines a target-aware Direction Oracle, a soft binding-affinity gate, and amortized fine-tuning of a pre-trained discrete diffusion model, enabling targeted agonist and antagonist generation decoupled from binding affinity and unattainable by equilibrium-based or inference-only guidance baselines. The code and checkpoints are available at https://huggingface.co/ChatterjeeLab/TD3B.

Many medicines work by switching protein receptors in our cells ON or OFF — blockbuster weight-loss drugs activate the GLP-1 receptor, while some sleep medications block orexin receptors. Today's AI tools for designing drug-like molecules mostly check only whether a candidate sticks to its target. They cannot reliably tell whether it will switch the receptor on (an "agonist") or off (an "antagonist"). Two molecules can latch onto the same pocket yet produce opposite biological effects, so "binds well" is a misleading goal. We built TD3B, an AI that designs short protein-like molecules called peptides with a chosen switching direction. Instead of asking only "does it bind?", TD3B asks "does it bind AND push the receptor the way I want?" It combines a direction predictor trained on known drug data, a filter that screens out non-binders, and a fine-tuning step that steers a generative model toward the desired effect. TD3B is the first sequence-based generative model that lets researchers explicitly request activator or blocker behavior for a chosen target. We demonstrate it on receptors relevant to obesity, diabetes, and sleep disorders. By separating "what binds" from "what activates," our approach could reduce a major cause of late-stage failure in drug discovery.