Diffusing protein binders to intrinsically disordered proteins

Abstract

Proteins that bind to intrinsically disordered proteins (IDPs) and intrinsically disordered regions (IDRs) with high affinity and specificity could be useful for therapeutic and diagnostic applications^1-4. However, a general methodology for targeting IDPs or IDRs has yet to be developed. Here we show that starting only from the target sequence of the input, and freely sampling both target and binding protein conformations, RFdiffusion⁵ can generate binders to IDPs and IDRs in a wide range of conformations. We used this approach to generate binders to the IDPs amylin, C-peptide, VP48 and BRCA1_ARATH in diverse conformations with a dissociation constant (K_d) ranging from 3 to 100 nM. For the IDRs G3BP1, common cytokine receptor γ-chain (IL-2RG) and prion protein, we diffused binders to β-strand conformations of the targets, obtaining K_d between 10 and 100 nM. Fluorescence imaging experiments show that the binders bind to their respective targets in cells. The G3BP1 binder disrupts stress granule formation in cells, and the amylin binder inhibits amyloid fibril formation and dissociates existing fibres, enables targeting of both monomeric and fibrillar amylin to lysosomes, and increases the sensitivity of mass spectrometry-based amylin detection. Our approach should be useful for creating binders to flexible IDPs or IDRs spanning a wide range of intrinsic conformational preferences.

Fig. 1. Design strategies for binding conformational flexible IDPs or IDRs.

A, Design approach. Aa, Disordered structures of the selected targets. Left, AF2 structure predictions for VP48, BRCA1, ARATH and FUS, coloured by pLDDT scores, and NMR structures of amylin (PDB ID: 2KB8) and CP (PDB ID: 1T0C). Targeted regions are shown as spheres. Right, diffusion models for proteins are trained to recover noised protein structures and to generate new structures by reversing the corruption process through iterative denoising of initially random noise into a realistic structure. Five representative trajectories are shown for amylin, which result in five different conformations of the target. Ab, Diffusion with constrained secondary structure. Left, AF2 predictions for G3BP1 and IL-2RG. Right, a modified version of RFdiffusion was trained, allowing for specification of the secondary structure of a region to be helix or strand conformations, along with the sequence. B, Two-sided partial diffusion. Varying the extent of initial noising (top row) enables control over the extent of the introduced structural variation (colours indicate new designs, and grey denotes the parent design).

Fig. 2. Design of intrinsically disordered region binders.

a–d, Computational design models of amylin and designed binders amylin-68n_αβ (a), amylin-36_αβ (b), amylin-75_αα (c) and amylin-22_αβL (d) generated using sequence input diffusion (top), with a zoom-in view of disulfide bonds formed during diffusion. The secondary structure of amylin is indicated by the binder name subscripts. BLI measurements of the designed binder–amylin interaction are shown below the design models. e–g, Design of binders for CP (e), VP48 (f) and BRCA1_ARATH (g) using sequence input diffusion. h–j, Design of binders for G3BP1 (h), prion protein (i) and IL-2RG (j) with strand specification during diffusion.

Fig. 3. Structural characterization.

a, Designed model of amylin-22_αβL in complex with amylin; amylin and designed binder are shown in dim grey and grey, respectively (left). Helical and strand segments forming the binding groove are highlighted by blue dashed ellipsoids. Right, crystal structure of the designed complex at 1.8 Å resolution, with the target and binder in blue and cornflower blue, respectively. b, Left, overlay of the design model and crystal structure of amylin-22_αβL. Right, zoom-in on the designed interface. Binder proteins are rendered with 90% transparency, and key amylin residues are labelled. c, Left, design model of G3BP1-11. Two α/β-topologies (T1 and T2) of the binders are labelled. The front helix of T2 is denoted by a black arrow. Right, the crystal structure of G3BP1-11 at 2.4 Å resolution, with the target and binder proteins rendered in dark red and rosy brown, respectively. d, Left, overlay of the design model and the crystal structure of G3BP1-11. Right, zoom-in on the designed interface. e, Heatmaps representing C-peptide-binding K_d (nM) values for single mutations in the designed interface (left), core (middle) and surface (right). Depleted substitutions are in blue, beneficial substitutions are in red, and grey indicates the lost yeast strains. Strand 1 (interface), the right segment of strand 2 (core) and the exposed surface residues without contacts (surface) are highlighted (Extended Data Fig. 7e). f, Designed binders coloured by positional Shannon entropy from site saturation mutagenesis (top; conserved in blue and variable in red), and zoomed-in views of the design interface and core with C-peptide (bottom).

Fig. 4. Designed binders bind to their targets in cells and binder G3BP1-11 modulates G3BP1 phase separation.

a, Colocalization of IL-2RG-30, G3BP1-11 and BRCA1_ARATH-35 with their full-length targets in HeLa cells. Scale bars, 10 µm (top row) and 20 µm (middle and bottom rows). b, Colocalization of amylin-22, amylin-36, amylin-68n, VP48-2 and CP-35 with their targets. Scale bars, 20 µm. c, The LC–MS/MS recovery percent of amylin from phosphate-buffered saline (PBS)–0.1% 3-((3-cholamidopropyl) dimethylammonio)-1-propanesulfonate (CHAPS) buffer and EDTA-anticoagulated plasma was compared between BSA-blocked tosyl-activated bead, an off-target binder, and the binder amylin-68n. Percent recovery was calculated relative to the peak area of pure amylin in elution solvent (100% recovery). The data indicate mean ± s.d. (n = 3 technical replicates). d,e, Phase diagrams showing effects of G3BP1-11 (d) and control (e) on G3BP1–RNA phase separation across 0–8 ng μl⁻¹ RNA. Red indicates phase-separated droplets, and blue indicates mixed states. f,g, Without binder, confocal images indicate that G3BP1 forms cytoplasmic puncta after 1 h of arsenite treatment (f). Quantification of G3BP1 puncta per cell (each point represents one cell; n = 20 cells per condition), showing a notable increase upon arsenite treatment, is also displayed (g). h,i, With the G3BP1-11 binder, arsenite treatment failed to induce a similar increase in puncta. Confocal images (h) show colocalization of G3BP1 (green) and the co-expressed mScarlet-tagged binder (red). Quantification revealed no markedly increase in puncta after arsenite treatment (i). Scale bars, 10 μm (f,h).

Fig. 5. Amylin fibril disruption and degradation using designed binder.

a,b, NS-EM visualization of fibril dissociation by amylin-36_αβ at both elongation (a) and mature (b) phases. Scale bars, 100 nm. c, The ThT assay revealed that all four binders could strongly inhibit fibril formation at a binder:amylin molar ratio of 1:4. d, Amylin-36_αβ could dissociate fibrils at the elongation phase in a concentration-dependent manner. Amylin-36_αβ was added at 3 h during the elongation phase of amylin fibrils (as marked by the dashed line), and the ThT assay was performed to monitor the process. The red and blue dots indicate amylin-36_αβ:amylin molar ratios of 1:4 (10 μM binder) and 1:40 (1 μM binder), respectively. e, The ThT assay was performed after the mature amylin fibrils were formed for 24 h and, at the same time, amylin-36_αβ was added. The red and blue dots indicate that the amylin-36_αβ:amylin molar ratios are 1:4 (10 μM binder) and 1:40 (1 μM binder), respectively. f,g, Confocal microscopy images of Hep3B cells treated with Alexa Fluor 647-labelled amylin fibrils (f; 24 h incubation at 37 °C) or monomers (g), in complex with amylin-36 (top), amylin-36–ASGPR (middle) or amylin-36–IGF2R (bottom). DAPI (blue) stains nuclei, LAMP1 (green) marks lysosomes, and amylin (magenta) is visualized via fluorophore-conjugated streptavidin. Composite panels (rightmost) show merged channels. Scale bars, 25 μm (f,g). h, Quantification of internalized amylin fibrils (left) and monomers (right) by flow cytometry; the median fluorescence intensity (MFI) was measured. Data are shown as mean ± s.d. from three biological replicates.

Extended Data Fig. 1. Diffusing de novo binder design to amylin.

a, Top, the designed structures of four initial hits, Amylin-1227, −4036, −4188, −562, which serve as starting points for two-sided partial diffusion. Bottom, bio-layer interferometry (BLI) result of the four hits revealing the binding affinity of the four initial hits are 100, 317, 431, 454 nM, respectively. b, Circular dichroism data show that the four optimized binders have helical secondary structure and are stable up to 95 °C (inset). c, The per residue pLDDT (predicted Local Distance Difference Test) plotting of Amylin- Amylin-22_αβL complex in design.

Extended Data Fig. 2. Two sided partial diffusion and comparison with one sided partial diffusion.

a, Top, two-sided partial diffusion allows simultaneous conformational changes in both the target and the binder. Bottom, one-sided partial diffusion solely diversifies the conformation of the binder while keeping the target fixed. b, Two-sided partial diffusion (in red) diversifies the target while one sided partial diffusion (in blue) keeps the target fixed. c, The target-binder complex diverse magnitudes of two-sided (in red) and one-sided partial diffusion (in blue) remain comparable before nosing step 35, after step 35, the diverse magnitude of two-sided partial diffusion is larger than one sided one. d, Take the interface pAE_interaction <10, pLDDT >90 as cutoff criterion, two-sided partial diffusion yielded designs with generally better metrics than one sided diffusion. At steps 25, 30, and 35 exclusively, one-sided partial diffusion exhibited superior performance. However, in practical cases, we typically operate within fewer than 25 steps to remain the main features of parent structure.

Extended Data Fig. 3. Diffusing de novo binder design to CP and disorder and secondary structure prediction for CP and VP48.

a, IUPred3 predictions for the CP, the predicted disorder scores remain above 0.5 across the targeted regions, indicating that CP is intrinsically disordered^,. b, JPred4 (ref. ) secondary structure predictions for CP, three residues within the target region show strand propensity (green arrows). c, Sequence-input diffusion was carried out, allowing CP to sample diverse conformations. Representative examples are shown here. The diverse conformations of CP and protein binder are rendered in blue and wheat colour, respectively. d, Design model of the initial hit CP-95 which was also the starting point of two-sided partial diffusion. e, The BLI data revealed that the binding affinity of the initial hit CP-95 is 16 μm. f, Scatter plot showing the distribution of designs based on the number of hydrogen bonds (hbond_number) and the RMSD of the binder (rmsd_binder). Each blue dot represents a design, while the red dot marks a validated hit. The dashed black lines indicate the cutoff values based on the initial hit criteria (hbond_number = 13 and rmsd_binder = 0.545). g and j, Circular dichroism data show that the binder CP-35 (g) and VP48-2 (j) have helical secondary structure and are stable up to 95 °C (inset). h, Jpred4 predictions for VP48, indicating VP48 doesn’t have secondary structure propensity. i, Predicted disorder profiles for VP48 generated by IUPred3 indicate low disorder propensity in its central segments. Combined with the low pLDDT scores predicted by AlphaFold (Fig. 1A), these results suggest that VP48 is structurally ambiguous.

Extended Data Fig. 4. Diffusing de novo binder design to BRCA1_ARATH and FUS.

a,e, IUPred3 predictions for the targeted regions on BRCA1_ARATH (a) and FUS (e); the predicted disorder scores remain above 0.5 across the targeted regions, indicating that they are intrinsically disordered. b, JPred4 secondary structure predictions for the targeted regions on BRCA1_ARATH; five residues within the target region show strand propensity (green arrows). c, Top, the designed structures of two initial hits of BRCA1_ARATH, YTE-19 and YTE-22, which serve as starting point of two-sided partial diffusion. Bottom, the BLI result of the two hits, revealing that the binding affinities of the two initial hits are 420 and 450 nM, respectively. d,h, Circular dichroism data show that the binder BRCA1_ARATH-35 (d) and FUS-40 (h) have helical secondary structure and are stable up to 95 °C (inset). f, No secondary structure is predicted within the targeted region for the targeted region on FUS predicted by Jpred4. g, Top, the designed structures of three hits of FUS, FUS-40, FUS-42 and FUS-47. Bottom, the BLI result reveals that the binding affinity of the three initial hits are 520, 730 and 750 nM, respectively.

Extended Data Fig. 5. Diffusing de novo binder design to G3BP1^RBD.

a, IUPred3 predictions for the C-terminus of G3BP1, with the targeted regions highlighted by black dashed boxes and the corresponding sequences shown in blue, indicating that the segment we targeted on G3bp1 is intrinsically disordered. b, JPred4 secondary structure predictions for the same regions on G3BP1; no secondary structure is predicted within the targeted region. c, Comparative analysis of structural outcomes between sequence input and strand specification approaches in protein design. The table presents the number of designs (10k) and the distribution of secondary structures (helix:strand:loop) for both methods. This table counts the successful cases where the pAE_interaction is less than 10 and the plddt_binder score is greater than 90, noting 23 successes with sequence input and 1,192 with strand specification. This reflects an approximately 51-fold increase in efficacy with the strand specification method, highlighting its superior performance in achieving desired structural configurations. d, The 23 successful cases designed using sequence input RFdiffusion all feature targets in strand conformation. e, Design models (top) and BLI data (bottom) of the four initial hits of G3BP1^RBD which was also the starting point of two-sided partial diffusion. f, Circular dichroism data show that the G3bp1-11 binder has helical secondary structure and is stable up to 95 °C (inset).

Extended Data Fig. 6. Diffusing de novo binder design to prion protein.

a, Jpred4 prediction for the prion target region, outlined by a dashed box, shows a β-strand prediction for five residues. b,g, Circular dichroism data show that the PRI28 binder (g) and IL-2RG-30 (g) has helical secondary structure and is stable up to 95 °C. c, The design model of PRI22, designed using target sequence information alone. d, The BLI data revealed that the binding affinity of PRI22 is 1.88 μM (left), which improved to 80 nM after two-sided partial diffusion (right). e, The specificity test for prion binder PRI28 (Fig. 2i) against various amyloid target sequences showed that PRI28 is highly specific, with some cross-reactivity observed only with TEME106B. f, Jpred4 secondary structure prediction for the IL-2RG target region, with the designed sequence highlighted by a black dashed box. Within this region, a short β-strand (green) is predicted for five residues, indicating localized strand propensity.

Extended Data Fig. 7. Crystal structure of Amylin-18_αβ and SSM analysis of CP-35.

a, Left, the designed model of Amylin-18_αβ, with target and binder proteins rendered in dim grey and grey, respectively. Right, the crystal structure of Amylin-18_αβ at 2.0 Å-resolution, with target and binder proteins rendered in salmon and tan, respectively. b, Left, the overlay of the design model and the crystal structure of Amylin-18_αβ. Right, magnified views of the regions indicated with blue dotted frames in the left panel. The crystal structure of Amylin-18_αβ closely recapitulates the design model, with a Ca RMSD 0.741 Å for entire complex between design and crystal structure. The interface residues are nearly perfectly aligned with the design model structure, showing interface Ca and sidechain RMSD of 0.958 and 1.279, respectively. c, The crystal structure of G3bp1-11, positioned 4 Å away from the target on the binder, is marked in blue. d, Full SSM maps for the design of CP-35. e, Zoomed-in views of the residues presented in the surface region, as shown in Fig. 3e.

Extended Data Fig. 8. Designed binders show high specificity for their targets.

Biotinylated peptides were immobilized onto octet streptavidin biosensors at equal densities and incubated with all binders in separate experiments at three concentrations. Amylin-68n_αβ, −36_αβ, −75_αα, −22_αβL are abbreviated as Am68n, Am36, Am75 and Am22, respectively. The designed on-target interactions are indicated with a light red background.

Extended Data Fig. 9. Point mutations at binder–target interfaces disrupt binding, while the designs inhibit Amylin fibril formation and dissociate existing fibrils.

a-c, Designed complex structures of binders (light pink) in complex with their respective IDP targets (slate) amylin (a, binder Amy36 as representative), C-peptide (b), and VP48 (c). Point mutations introduced at the predicted interface are highlighted with arrows and dashed circles: L12T and L16T for amylin; L26T and L30T for C-peptide; and M10K and L24T for VP48. d, Amylin binders Amylin-22_αβL and Amylin-36_αβ inhibit fibril formation in a concentration-dependent manner. The initial concentration of Amylin monomer was 10 μM, with subsequent additions of binders at 2.5 μM, 0.25 μM, and 0.025 μM, establishing molar ratios of binder to Amylin of 1:4, 1:40, and 1:400, respectively. e,f, Negative stain electron microscopy images were taken of 40 μM Amylin monomer samples following the addition of 10 μM Amylin-36_αβ (e) and Amylin-22_αβL (f) at 1 h and 18 h, respectively. Scale bars, 100 nm.

Extended Data Fig. 10. The effect of the binder on G3bp1 phase separation at higher polyA RNA concentrations.

Phase diagrams comparing the effects of the G3BP1-11 binder (left) and control (right) on phase separation between G3BP1 and RNA over an RNA concentration range of 0–150 ng/μl.