RFdiffusion Opens a New Route to Binders for Intrinsically Disordered Proteins

Overview

Intrinsically disordered proteins (IDPs) and intrinsically disordered regions (IDRs) do not remain in one stable three-dimensional structure. Their conformational flexibility is central to signalling, phase separation and molecular recognition, but it also makes them difficult targets for conventional structure-based design.

In a 2025 Nature paper, Liu and colleagues reported a general RFdiffusion-based strategy for designing folded protein binders from the target amino-acid sequence while allowing the target conformation and binder structure to emerge together. The study produced nanomolar binders to diverse disordered targets, validated selected complexes structurally and demonstrated functional effects in cells and in amyloid-related assays.

Why Disordered Proteins Are Hard to Target

Many drug-design methods assume that the target presents a reproducible pocket or interface. IDPs instead populate ensembles of rapidly interconverting conformations. A structural snapshot may capture only one environment-dependent state, and a pocket observed in that state may not be representative of the ensemble.

Disorder is not equivalent to lack of function. Flexible proteins frequently use short motifs, transient secondary structure and context-dependent interactions to regulate cellular processes. Their adaptability can enable one region to interact with multiple partners, but it also creates three practical design problems:

• No single receptor geometry: fixing the target in an arbitrary conformation can bias the design toward a state that is rarely populated.
• Limited interaction surface: short disordered segments may not provide enough pre-organized shape for conventional docking.
• Specificity risk: a binder that mainly recognizes a common helix or β-strand pattern may cross-react with unrelated proteins that adopt similar local structure.

The authors note that IDPs and proteins containing IDRs represent a substantial fraction of the human proteome. This makes a general binder-design method potentially valuable for research reagents, biomarker capture and future therapeutic engineering, provided specificity and biological delivery can be controlled.

Sequence-Input Diffusion and Conformational Co-Sampling

The design strategy adapts RFdiffusion so that target backbone coordinates are not required as a fixed input. The target sequence is provided while part or all of its structure is noised. During iterative denoising, both target and binder conformations are sampled. The resulting complex is then sequence-designed and filtered for structural confidence and interface quality.

Two-sided partial diffusion

Initial amylin designs bound in the 100–454 nM range. To improve affinity, the team used two-sided partial diffusion, in which both the binder and target are partially noised and jointly denoised. Allowing the target to adapt while preserving information from a parent complex improved shape complementarity. Among 174 top-ranked optimized amylin designs, 107 bound the target; the best reported binders covered distinct target conformations with dissociation constants from 3.8 to 100 nM.

Secondary-structure specification

For shorter regions with β-strand propensity, the authors modified the model so that the target sequence could be paired with a desired local secondary structure without prescribing full three-dimensional coordinates. This strategy was used for G3BP1, IL-2RG and prion protein segments and improved success for challenging short IDRs.

A Broad Panel of Disordered Targets

The study tested targets with different lengths, sequence properties and biological roles. Sequence-input diffusion generated binders for amylin, C-peptide, VP48 and a BRCA1-associated construct in diverse conformations. Secondary-structure-specified diffusion was used for G3BP1, IL-2RG and prion protein regions.

Structural and Specificity Validation

Computational design becomes compelling only when the intended binding mode is experimentally supported. The team solved crystal structures for selected amylin and G3BP1 complexes. The G3BP1-11 complex showed a Cα RMSD of approximately 0.8 Å between the designed and experimental complex, while the amylin-22 complex reproduced the designed groove and interface geometry at high resolution.

When a C-peptide complex could not be crystallized, site-saturation mutagenesis provided a lower-resolution footprint: substitutions at the predicted interface and binder core were less tolerated, supporting the model. An all-by-all binding panel showed strong target specificity overall, with weak cross-reactivity observed only at high concentrations in two tested cases.

From Binding to Functional Control

Cellular target engagement

Fluorescence colocalization experiments showed that multiple designed proteins could fold and engage their corresponding full-length targets in mammalian cells. Interface mutations disrupted colocalization for selected short targets, providing additional evidence that recognition depended on the designed contact surface.

Regulating G3BP1 phase separation

G3BP1 is a central regulator of stress-granule assembly. The G3BP1-11 binder shifted the G3BP1–RNA phase boundary in vitro and reduced arsenite-induced puncta formation in cells. This result illustrates how an IDR binder can serve as a mechanistic tool, not only a capture reagent.

Inhibiting and reversing amylin fibrils

At a 1:4 binder-to-amylin molar ratio, the tested binders strongly inhibited fibril formation. Amylin-36 also dissociated pre-formed fibrils in a concentration-dependent manner. When fused to receptor-targeting EndoTags, the binder promoted cellular uptake and lysosomal trafficking of monomeric and fibrillar amylin. These experiments demonstrate functional potential, but they do not establish therapeutic efficacy in animals or patients.

Why This Work Matters

The key advance is methodological. Instead of choosing one uncertain structure for a disordered target, the model searches target and binder conformations together. This reframes disorder from an obstacle that must be removed into a design variable that can be sampled.

The approach could support several research directions:

• conformation-selective probes for dissecting IDP biology;
• reagents for biomarker enrichment and detection;
• binders that modulate phase separation or transient protein interactions;
• starting points for targeted localization or degradation strategies; and
• new ways to interrogate proteins that lack persistent small-molecule pockets.

These opportunities should be separated from therapeutic readiness. Designed binders must still be optimized for specificity, stability, expression, delivery, immunogenicity and performance in disease-relevant models.

Limitations and Practical Design Risks

• Large computational funnels: approximately 10,000–50,000 diffused designs were generated for each target before filtering to experimentally manageable sets.
• Optimization is often required: initial hits may need partial diffusion or additional sequence refinement to reach strong affinity.
• Conformation-dependent specificity: binders may recognize a local secondary-structure pattern shared by off-target proteins.
• Structural validation is not always available: crystallization can fail, requiring mutational or biophysical alternatives.
• Biological context matters: intracellular localization, post-translational modification, partner competition and target concentration can change performance.
• Therapeutic translation remains unproven: the reported results are research-stage demonstrations, not evidence of clinical safety or efficacy.