GCSRdata

De Novo protein binder design is pivotal for biomedical and biotechnological applications. Although
various diffusion-based computational design methods have emerged recently, they are still hindered
by low success rates. Consequently, extensive high-throughput screening is often required to identify
viable candidates, leaving substantial room to improve overall design efficiency. By analyzing the 3D
structural evolution of sampling trajectories, we hypothesize that these limitations primarily stem from
monotonic denoising traps, wherein the generative model is incapable of correcting early suboptimal
decisions. Consequently, we present Geometrically Constrained Stochastic Resampling (GCSR),
a plug-and-play strategy that enhances RFdiffusion without altering its core architecture. GCSR
introduces hierarchical dynamic sampling, effectively disrupting monotonic trajectory drift through
time jumps and resampling operations. Furthermore, we design an SE(3)-decoupled perturbation
mechanism governed by trust-region constraints. This mechanism precisely drives the binder toward
superior target surface while rigorously preserving the structural integrity of the protein backbone.
Experimental in silico evaluations across five therapeutically significant targets with diverse interfacial
characteristics (InsulinR, PD-L1, IL-7R𝛼, EGFR, and TrkA) demonstrate statistically significant
improvements in in silico success rates. Furthermore, comprehensive ablation studies on the EGFR
target validate the organic synergistic mechanism underlying GCSR: Jump-perturbation resampling
strategy synergizing coarse and fine granularities, SE(3)-decoupled perturbation, and biophysical
trust-region constraints.