COMPASS: Decoupled Latent Steering for Protein Conformational Transitions

Protein function often depends on distinct conformational states and the transitions between them, yet structure predictor models such as AlphaFold 3 (AF3) are not designed to capture all biologically meaningful states. Unsteered ensemble generation models can broadly explore conformational space but lack directional control, whereas direct 3D-coordinate guidance accelerates structural transitions at the cost of geometric violations and steric clashes. We propose COMPASS, a framework that decouples target-directed optimization from structural inference. COMPASS backpropagates gradients only to conditional prior embeddings at an intermediate anchor timestep via Tweedie's expectation, leaving the reverse diffusion process unmodified. On a curated set of paired ligand-bound and unbound protein structures, COMPASS improves the efficiency of sampling structures that closely resemble a target conformational state by up to 18.9% over the unsteered baseline without compromising the physical quality of the structure. Notably, Apo-guided steering raises the Holo-like hit rate from 20.7% to 50.0%, suggesting that COMPASS broadens sampling across functionally distinct regions of the transition manifold rather than simply steering generations toward the guide structure itself. Downstream ligand docking on COMPASS steered structures, evaluated on standard and ternary-complex benchmarks, shows docking accuracy comparable to AF3 protein-ligand complex predictions.