Go to ICML 2026 Workshop GenBio homepageAlloGen: Conformation-Selective Binder Design with Differential State Scoring
Keywords: Protein binder design; Conformational Selectivity; Allosteric proteins; Diffusion model
Abstract: Protein binder design has largely optimized for affinity alone, leaving conformational selectivity—stabilizing one functional state while disfavoring others—unaddressed. This gap is consequential for allosteric targets such as kinase activation loops, nuclear receptors, and GPCRs, where selective state engagement, not affinity alone, determines therapeutic utility.
We introduce AlloGen, the first framework that explicitly incorporates conformational selectivity into protein binder design. AlloGen decouples generation from selectivity evaluation through a learned scorer $Q_\theta$: a lightweight SE(3)-invariant interface graph transformer trained via a two-phase curriculum. Phase 1 grounds $Q_\theta$ in interface geometry through DockQ regression; Phase 2 imposes conformational discrimination through contrastive InfoNCE fine-tuning on paired two-state conformations. The frozen $Q_\theta$ is generator-agnostic and integrates with any backbone generator—including RFdiffusion, PXDesign, and Proteina-ComplexA—as either a passive reranker or an active gradient-based guide, with no retraining required.
Trained on 65 targets across 15 protein families, $Q_\theta$ generalizes to held-out out-of-distribution targets ($\bar{\rho} = 0.520$) where energy-based baselines fail entirely, demonstrating that it captures genuine conformational preference rather than trivial structural signals. Extended to guided generation, all evaluated strategies achieve positive selectivity, with RFdiffusion + Langevin refinement reaching the highest selectivity ($\bar{S} = +0.677$).
Email Sharing: We authorize the sharing of all author emails with Program Chairs.
Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.
Submission Number: 21
Loading