Proteins often work by switching between states. Small molecules or peptides called ligands can bias the direction of those switches. For example, an agonist pushes a receptor toward an active state, while an antagonist prevents activation. Most current design methods try to stabilize one protein shape. They miss the more subtle task of steering the direction of state changes. This paper introduces TD3B (Transition-Directed Discrete Diffusion for Allosteric Binder design), a sequence-based method intended to design binders with a specified agonist or antagonist effect.

The authors build TD3B on large, pre-trained sequence models called masked discrete diffusion language models (MDLMs). Rather than retraining such a model from scratch, they add a directional guidance layer. TD3B combines three components: a Direction Oracle that predicts whether a candidate binder will act like an agonist or antagonist for a given target; a soft binding-affinity gate that favors sequences likely to bind without making affinity the only goal; and an amortized fine-tuning scheme that steers the MDLM toward sequences that match the desired direction. The team frames the problem as learning a sequence-conditioned transition operator over protein macrostates, which makes directionality an explicit design objective.

At a high level TD3B works by biasing the pre-trained sequence model toward a reward-tilted distribution of peptide sequences. The fine-tuning uses importance-weighted denoising and a contrastive loss to separate agonist-like from antagonist-like representations. The Direction Oracle provides a binary signal (+1 or −1) indicating the desired sign of transition asymmetry. The affinity predictor acts as a soft gate, so the model can prefer binders that both bind reasonably and push the transition in the chosen direction. The authors emphasize they are not trying to predict exact kinetic rates. Instead, they use a coarse, learnable supervision that captures the sign of the directional effect.