AlloGen: Conformation-Selective Protein Binder Design

• AlloGen is a computational framework that designs protein binders with precise conformational selectivity by decoupling structure generation from state-specific discrimination.
• It employs an SE(3)-invariant interface graph transformer and a differentiable scoring function to distinguish between desired and undesired receptor states.
• The framework integrates passive reranking and active gradient-based guidance, validated by experimental binding assays on targets like calmodulin.

AlloGen is a modular computational framework designed to generate conformation-selective protein binders by decoupling structure generation from state-specific discrimination. It specifically addresses the longstanding limitation in protein binder design workflows which historically focus on optimizing binding affinity alone, neglecting the functional necessity for conformational selectivity in allosteric targets such as kinases, nuclear receptors, and GPCRs. AlloGen systematically incorporates conformational selectivity as a learnable and model-agnostic property by leveraging a differentiable scoring function that can operate alongside any generative backbone model without retraining, enabling both passive and active strategies for the design of state-specific molecular interactions (Cao et al., 3 Jun 2026).

1. Conformational Selectivity: Scientific Rationale

Functional specificity in protein recognition often depends not solely on affinity but on a binder’s ability to discriminate among the dynamic conformational states adopted by a receptor. In allosteric regulation and signal transduction (e.g., kinases alternately assuming DFG-in/out states, helix 12 movements in nuclear receptors, active/inactive cycling in GPCRs), engaging the desired signaling state while rejecting alternatives is essential to modulate function purposively. Conventional structure-based design pipelines are affinity-optimized against single static receptor structures and lack a formal mechanism to encode conformational selectivity. This leads to binders with high affinity but no functional discrimination, potentially stabilizing undesired signaling states or failing to achieve therapeutic or engineering objectives. AlloGen addresses this gap by explicitly optimizing for differential binding across structurally characterized states (Cao et al., 3 Jun 2026).

2. Architecture: Interface Graph Transformer and $Q_\theta$ Scorer

AlloGen’s core is the differentiable selectivity scorer $Q_\theta$, realized as an SE(3)-invariant interface graph transformer. For a candidate binder backbone $Y$ and two receptor conformations $X^0$ (undesired/apo) and $X^1$ (desired/holo), $Q_\theta$ returns scalar values in $(0,1)$, and the selectivity margin $\Delta Q$ is defined as:

$\Delta Q = Q_\theta(X^1, Y) - Q_\theta(X^0, Y)$

The interface graph $\mathcal G = (\mathcal V, \mathcal E)$ is constructed by including all residues within 8 Å (C$Q_\theta$0–C$Q_\theta$1 distances) across binder and receptor chains. Nodes are featurized by amino acid identity, backbone dihedrals $Q_\theta$2, $Q_\theta$3, $Q_\theta$4, sidechain torsions $Q_\theta$5, $Q_\theta$6, and optional ESM-2 embeddings, while edges encode SE(3)-invariant geometric features (distance, relative orientation, sequence separation, chain identity). The model employs a four-layer edge-biased graph transformer, with multi-head edge-conditioned attention and SE(3)-invariant pooling strategies to produce state scores via an MLP and sigmoid activation.

3. Two-Phase Curriculum Training

Training $Q_\theta$7 to assess both interface quality and conformational discrimination is accomplished via a two-phase curriculum:

• Phase 1: Geometry Regression $Q_\theta$8 is trained on a diverse dataset of holo native complexes, apo-mismatches, rigid-body/FastRelax decoys, and generative negatives. The score is regressed to the DockQ interface quality proxy using mean squared error:

$Q_\theta$9

• Phase 2: Contrastive State Discrimination With Phase 1 weights fixed, paired triplets $Y$0 from the same receptor are used to enforce selectivity via multi-negative InfoNCE loss:

$Y$1

Binder-side dropout is applied (probability 0.3) to enforce geometry reliance at inference.

4. Generator-Agnostic Integration and Guidance Modes

AlloGen's main operational advantage is its generator-agnostic nature. Once $Y$2 is trained, it can be interfaced with any fixed backbone generator $Y$3 in two principal ways:

Integration Mode	Mechanism	Effect
Passive reranking	Score $Y$4 samples, select highest $Y$5	No retraining; boosts selectivity
Active gradient-based guidance	Backpropagate $Y$6 during gen.	Enables real-time steering

Passive reranking operates by computing the logit-space selectivity margin $Y$7 for each sampled candidate and selecting maximizers. Reranking (e.g., best-of-5/10) substantially amplifies mean selectivity even with vanilla backbone generation.

Active guidance injects selectivity gradients at test time:

• Langevin refinement: Gradient ascent on $Y$8 applied to denoised backbone coordinates.
• Classifier guidance: Gradient term added in generative denoising steps.
• Twisted diffusion sampling (TDS) and Sequential Monte Carlo (SMC): Sampling or resampling strategies weighted by $Y$9 to enrich high-selectivity candidates.

Importantly, $X^0$0 operates entirely independently of generator parameters, so backbone generators such as RFdiffusion, PXDesign, Proteina-ComplexA, and sequence-level diffusion models require no retraining for state selectivity.

5. Benchmark Evaluation and Quantitative Performance

AlloGen was assessed on a two-state binder benchmark comprising 65 targets and 2,896 curated apo/holo receptor–binder complexes spanning 15 families (kinases, GTPases, nuclear receptors, GPCRs/ion channels, proteases, epigenetic readers, among others). On eight held-out out-of-distribution targets (including CaM, BCL-2, MDM2, ERα, A$X^0$1R, Ran, PAI-1, Integrin), $X^0$2 achieved mean Spearman $X^0$3 to the DockQ quality metric, outperforming classical heuristics like PRODIGY ($X^0$4) and interface size ($X^0$5), with genuine conformational discrimination on 7/8 targets.

AlloGen delivered positive mean selectivity margins in all 15 generator/guidance settings tested. For CaM, RFdiffusion+Langevin achieved $X^0$6 with an 88% success rate (states that are both designable and selective), and best-of-10 reranking reached $X^0$7. Selectivity gains varied with baseline generator performance—some targets benefitted primarily from passive strategies, while others required active guidance to escape low or near-zero selectivity.

6. Experimental Validation: Calmodulin Case Study

AlloGen's computational predictions were validated by experimental synthesis and binding assays on de novo 20–25 residue peptides designed to target the Ca$X^0$8-bound (holo) form of calmodulin (CaM):

• Top eight high-selectivity ($X^0$9) peptides, one low-$X^1$0 negative control, and a canonical M13 positive control were synthesized.
• Binding was assessed via bio-layer interferometry (BLI) for holo (with CaCl$X^1$1) vs apo (with EGTA) CaM.
• Five out of eight high-$X^1$2 designs bound holo CaM (with $X^1$3 from 46.6 nM to 1.06 μM) and exhibited no detectable binding to apo CaM.
• The low-$X^1$4 negative control did not bind either state.

These results confirm that $X^1$5’s computational selectivity margin correlates with experimentally observed conformational specificity.

7. Implications and Prospective Developments

AlloGen demonstrates that conformational selectivity is a learnable, transferable property at the protein–binder interface level. Its decoupling of a small, differentiable state scorer from large, static generative models enables modular augmentation of existing workflows for the design of allosteric drugs, biosensors, and synthetic molecular switches. Potential extensions include:

• Generalization to multi-state landscapes (beyond two-state systems)
• Direct integration into sequence-generation models for end-to-end design
• Application to highly conformationally dynamic therapeutic targets

The AlloGen model and codebase are publicly released, providing a resource for the design of conformation-selective binders across the proteome (Cao et al., 3 Jun 2026).