Main

Peptide hormones, such as parathyroid hormone (PTH), neuropeptide Y (NPY), glucagon (GCG) and secretin (SCT), which adopt α-helical structures on binding their receptors1,2,3,4, play key roles in human biology and are well-established biomarkers in clinical care and biomedical research (Fig. 1a). There is considerable interest in their sensitive and specific quantification9, which at present relies on antibodies that require substantial resources to generate, can be difficult to produce with high affinity, and often have less-than-desirable stability and reproducibility10,11,12,13,14. The loop-mediated interaction surfaces of antibodies are not particularly well suited to high-specificity binding of extended helical peptides—almost all anti-peptide antibodies bind their targets in non-helical conformations15. Designed proteins with very high stability can be readily produced with high yield and low cost in Escherichia coli; however, although there have been considerable advances in de novo design of binders for folded proteins6,7, the design of proteins that bind helical peptides with high affinity and specificity remains an outstanding challenge. Design of peptide-binding proteins is challenging for two reasons. First, proteins designed to bind folded proteins, such as picomolar-affinity hyper-stable 50–65-residue minibinders7,16, have shapes suitable for binding rigid concave targets, but not for cradling extended peptides. Helical peptides can readily associate to form coiled-coil assemblies, and this principle has been used to design binders for a calmodulin peptide17, but coiled-coil subunits generally self-associate in the absence of binding partners owing to considerable exposed hydrophobic surface, considerably reducing the effective target-binding affinity. Second, peptides have fewer residues to interact with, and are often partially or entirely unstructured in isolation18. Hence, there can be an entropic cost of structuring the peptide into a specific conformation19, which compromises the favourable free energy of association. Progress has been made in designing peptides that bind to extended β-strand structures20 and polyproline II conformations21 using protein side chains to interact with the peptide backbone, but such interactions cannot be made with α-helical peptides owing to the extensive internal backbone–backbone hydrogen bonding.

Fig. 1: Design strategies for binding helical peptides.
figure 1

a, Helical peptide targets: apoptosis-related BH3 domains of Bid43 (PDB ID: 4QVE) and Bim44 (PDB ID: 3FDL), GCG3 (PDB ID: 1GCN), gastric inhibitory peptide45 (GIP; PDB ID: 2QKH), SCT46 (PDB ID: 6WZG), GCG-like peptide 147 (GLP1; PDB ID: 6X18), PTH48 (PDB ID: 1ET1), PTHrp25 (PDB ID: 7VVJ), PYY49 (PDB ID: 2DEZ) and NPY50 (PDB ID: 7X9A). b, Parametric approach. Left: sampling groove scaffolds varying supercoiling and helix distance to fit different targets. Middle: design model (spectrum) and PTH target (purple) of the best parametrically designed PTH binder. Right: split NanoBiT titration of PTH and the binder showed weak binding. a.u., arbitrary units. c, Inpainting binder optimization. Left: redesign of parametrically generated binder designs using RFjoint Inpainting to expand the binding interface and ProteinMPNN to redesign the sequences. Middle: AF2 prediction of Inpainted design (spectrum) with extended interface (teal), and PTH target (purple). Right: FP measurements (n = 4) indicate 6.04 nM binding to PTH and weak binding to off-target PTHrp. d, Threading approach to peptide binder design. Left: starting with a helix-bound scaffold, a target is threaded onto the bound helix and the interface is redesigned. Middle: AF2 prediction of design (spectrum) and SCT target (orange). Right: FP measurements (n = 4) indicate 3.95 nM binding to SCT and 12 nM binding to GCG. e, Hallucinating peptide binders. Left: Markov chain Monte Carlo (MCMC) steps are carried out in sequence space. At each step, the peptide sequence is re-predicted, and changes are accepted or rejected on the basis of interfacial contacts and AF2 metrics. The final structure is then redesigned using ProteinMPNN to avoid adversarial sequences. Interaction pAE, predicted alignment error across the interface; pLDDT, predicted local distance difference test. Middle: AF2 prediction of design (spectrum) and Bid target (blue). Right: FP measurements (n = 4) indicate 7 nM binding affinity to Bid.

Full size image

Design of peptide-binding scaffolds

We set out to develop general methods for designing proteins that bind peptides in helical conformations. To fully leverage recent advances in protein design, we explored both parametric and deep learning-based approaches. For parametric generation, we reasoned that helical bundle scaffolds with an open groove for a helical peptide could provide a general solution to the helical peptide-binding problem: the extended interaction surface between the full length of the helical peptide target and the contacting helices on the designed scaffold could enable high-affinity and specific binding, and the helices flanking the groove could limit self-association of the recessed hydrophobic surfaces. In parallel, we reasoned that deep learning methods, which do not pre-specify scaffold geometries, could permit the exploration of different potential solutions to peptide binding.

We began by exploring parametric methods for generating backbones with overall ‘groove’ shapes. Using the Crick parameterization of α-helical coiled coils22, we devised a method to sample scaffolds consisting of a three-helix groove supported by two buttressing helices (Fig. 1b and Methods). We assembled a library sampling a range of supercoiling and helix–helix spacings to accommodate a variety of helical peptide targets (Supplementary Figs. 1–3). We then used this library to design binders to PTH, GCG and NPY, and screened 12 designs for each target using a NanoBiT split-luciferase binding assay (Supplementary Fig. 4). Many of the designs bound their targets (3, 4 and 8 out of 12 to PTH, GCG and NPY, respectively) but with only micromolar affinities (Fig. 1b and Supplementary Fig. 4a–c). These results suggest that groove-shaped scaffolds can be designed to bind helical peptides, but also that design method improvement was necessary to achieve high-affinity binding.

We next explored using RoseTTAFold Inpainting (RFjoint)23, a model that can jointly design protein sequences and structures, along with ProteinMPNN24, an improved sequence design method, to improve the modest affinity of our tightest parametrically designed PTH binder (Fig. 1c, left). We used RFjoint Inpainting to extend the binder interfaces and ProteinMPNN to redesign the sequences, reasoning that the combination of these two methods could lead to more favourable interactions with the peptide. Out of 192 designs tested, 44 showed binding against PTH in initial yeast display screening. Following size-exclusion chromatography (SEC), the best binder was found to bind with 6.04 nM affinity to PTH using fluorescence polarization (FP). Binding was specific: very little binding was observed to PTH-related peptide (PTHrp), a related peptide sequence with 34% sequence identity that binds the same receptor as PTH25 (Fig. 1c, right). Overall, the affinity of the starting PTH binder was improved by approximately three orders of magnitude, and the computational model of the highest-affinity binder had 19% greater surface area contacting the target peptide (the structural extension was critical to the improvement in binding affinity; sequence redesign with ProteinMPNN of the original binding interface did not measurably increase affinity; Supplementary Fig. 5). We used the same design strategy to generate higher-affinity binders for NPY and GCG. Using weak parametric binders as a starting point, we extended their binding interfaces and redesigned their sequences to generate a 231-nM-affinity binder for GCG and a 3.5-µM binder for NPY after screening 96 designs (Extended Data Fig. 1a,b).

As an alternative to de novo parametric design of scaffolds that contain grooves, we explored the threading of helical peptides of interest onto already existing designed scaffolds with interfaces that make extensive interactions with helical peptides26,27 (Fig. 1d, left, and Methods). We threaded sequences of peptides of interest onto these complexes, and filtered for interfacial hydrophobic interactions between the target sequence and the scaffolds17,26. The selected scaffolds were then redesigned in the presence of the threaded target sequence with ProteinMPNN24 and the complex was predicted with AlphaFold228 (AF2; with initial guess6) and filtered on AF2 and Rosetta metrics. Initial screening using yeast surface display identified 4/66 binders for SCT, which were expressed in E. coli. After purification, all four of the designs were found to bind with submicromolar affinity using FP, with the highest-affinity design binding with an affinity of 2.7 nM for SCT (Fig. 1d, right); we also made designs with a dissociation constant (Kd) of <100 nM to GCG-like peptide 1 and gastric inhibitory polypeptide (Extended Data Fig. 2a,b). The SCT binder design bound GCG, which has 44% sequence identity to SCT4,29, with fourfold weaker affinity than SCT (Fig. 1d, right).

Designing binders using Hallucination

We next explored the use of deep learning Hallucination methods to generate helical peptide binders completely de novo, with no pre-specification of the binder or peptide geometry (Fig. 1e, left inset, and Supplementary Fig. 6a). Hallucination or ‘activation maximization’ approaches start from a network that predicts protein structure from sequence, and carry out an optimization in sequence space for sequences that fold to structures with desired properties. This approach has been used to generate new monomers30, functional-site scaffolds23 and cyclic oligomers31. Hallucination using AF2 or RoseTTAFold has the advantage that neither the binder nor the peptide structure needs to be specified during the design process, enabling the design of binders to peptides in different conformations (this is useful given the unstructured nature of many peptides in solution; disordered peptides can bind in different conformations to different binding partners18). Hallucination directly optimizes metrics correlated with binding, albeit with the possible hazard of generating adversarial protein sequences31. We began by designing binders to the apoptosis-related BH3 domain of Bid (Fig. 1a). The Bid peptide is unstructured in isolation, but adopts an α-helix on binding to Bcl-2 family members32,33; it is therefore a model candidate for the design of helix-binding proteins. Starting from only the Bid primary sequence, and a random seed binder sequence (of length 60, 70, 80, 90 or 100 residues), we carried out a Monte Carlo search in sequence space, optimizing for confident binding to the target peptide (AF2 predicted local distance difference test, pLDDT; and predicted alignment error, pAE)6. The trajectories typically converged in 5,000 steps (sequence substitutions; Supplementary Fig. 6b), and the output binder structure was subsequently redesigned with ProteinMPNN, as previously described31. All designed binders were predicted to bind to Bid in a predominantly helical conformation; the exact conformations differ between designs because only the amino acid sequence of the target is specified in advance. This protocol effectively carries out flexible backbone protein design, which can be a challenge for traditional Rosetta-based design approaches for which deep conformational sampling can be very compute intensive. In line with our prediction that ‘groove’ scaffolds would offer an ideal topology for helical peptide binding, many of the binders from this approach contain a well-defined ‘groove’, with the peptide predicted to make extensive interactions with the binder, typically helix–helix interactions (Extended Data Fig. 3a).

We experimentally tested 46 of the Hallucinated designs (Extended Data Fig. 3a) by co-expression of a GFP-tagged Bid peptide and His-tagged binders, with co-elution of GFP and binder used as a readout for binding. Four of these designs were further characterized, and showed soluble, monomeric expression even in the absence of peptide co-expression (Extended Data Fig. 3b), and could be pulled down using Bid BH3 peptide immobilized on beads (Extended Data Fig. 3c). Circular dichroism experiments indicated that the Bid peptide was unstructured in solution, and that helicity increased on interaction with the Hallucinated proteins, in line with the design prediction (Extended Data Fig. 3d). The binders were highly thermostable and, unlike the native Bcl-2 protein Mcl-1, readily refolded after (partial) thermal denaturation at 95 °C (Extended Data Fig. 3e). FP measurements revealed a 7-nM-affinity binder to Bid peptide (Fig. 1e, right inset), a higher-affinity interaction than that between Bid and the native partner Mcl-1 (Extended Data Fig. 3f,g).

Design refinement with RFdiffusion

We next explored using the RoseTTAFold-based denoising diffusion method RFdiffusion8. RFdiffusion directly generates protein structures with diverse topologies, and is much more compute efficient than Hallucination. We first extended RFdiffusion to enable optimization of existing helical peptide-binders.

A long-standing challenge in protein design is to increase the activity of an input native protein or designed protein by exploring the space of plausible closely related conformations for those with predicted higher activity34. This is difficult for traditional design methods as extensive full-atom calculations are needed for each sample around a starting structure (using molecular dynamics simulation or Rosetta full-atom relaxation methods), and it is not straightforward to optimize for higher binding affinity without detailed modelling of the binder–target side-chain interactions. We reasoned that, by contrast, RFdiffusion might be able to rapidly generate plausible backbones in the vicinity of a target structure, increasing the extent and quality of interaction with the target guided by the extensive knowledge of protein structure inherent in RoseTTAFold. Typically, during the reverse diffusion (generative) process, RFdiffusion takes random Gaussian noise as input, and iteratively refines this to a new protein structure over many (T) steps (generally 200). Part way through this denoising process, the evolving structure no longer resembles ‘pure noise’, instead resembling a ‘noisy’ version of the final structure. We therefore reasoned that ensembles of structure with varying extents of deviation from an input structure could be generated by partially noising initial starting structures to different extents (for example, time step 70), and then denoising to a similar, but not identical, final structure (Fig. 2a; in this case, the input coordinates to RFdiffusion at time step 70 are from a noised starting structure, rather than a partially denoised random distribution).

Fig. 2: Peptide binder optimization with RFdiffusion.
figure 2

a, Top: partial diffusion. RFdiffusion is used to denoise a randomly noised starting design (left); varying the extent of initial noising (middle row) enables control over the extent of introduced structural variation (bottom row; colours, new designs; grey, original design). Bottom left: partial diffusion diversifies designs. Note that the greater the amount of noise added, the more dissimilar the outputs are to the starting structure. Bottom right: depiction of the helix binder optimization strategy. b, Top: design model (spectrum) of the partially diffused binder to NPY (green) and FP measurements (n = 4) indicating a 5.3 nM binding affinity to NPY target and selectivity over PYY (brown). ND, not detectable. Bottom: design model (spectrum) of the partially diffused binder to GCG (yellow) and FP measurements (n = 4) indicating a subnanomolar binding affinity to GCG and selectivity over SCT (orange). c, Left: model (spectrum with GCG in grey) aligns with 0.72 Å RMSD to the 1.95-Å crystal structure (teal and yellow) of the RFjoint Inpainted GCG binder. Right: model (spectrum with GCG in grey) aligns with 0.6 Å RMSD to the 1.81-Å crystal structure (teal and yellow) of the partially diffused GCG binder. d, Left: the crystal structures of the Inpainted (grey) and partially diffused (teal and yellow) GCG binders have considerable topological similarity; there are many small readjustments. Right: FP titrations (n = 4) with GCG indicate much tighter binding following partial diffusion. e, Left inset: the crystal structure of the partially diffused backbone (teal) shows how the newly introduced Ile13 increases shape complementarity compared to the phenylalanine in the Inpainted binder (crystal structure in grey; structures aligned on residues 16–29 of GCG). Middle: crystal structure of the partially diffused GCG binder (teal and yellow). Right inset: the backbone shifts in the partially diffused structure (teal) enable Tyr16 to make packing and hydrogen-bonding interactions with the peptide; Ser16 in the original design did not make any peptide contacts (grey).

Full size image

We implemented this ‘partial diffusion’ approach (Methods), and sought first to assess the extent to which protein structures could be resampled and refined with partial diffusion. As expected, partial diffusion allowed diversification of a starting protein fold, and the magnitude of this diversity could be tuned by varying how much noise was added to a starting structure (Fig. 2a). We next explored the ability of partial diffusion to ‘regularize’ native protein backbones using as a metric AF2 structure prediction from a single sequence. We found that RFdiffusion improves the ‘designability’ of protein backbones: ProteinMPNN sequence design on partially diffused native backbones (with high similarity to the native fold, Extended Data Fig. 4a,c, middle row) have improved structure recapitulation (self-consistency) by AF2 compared to both the native sequence (Extended Data Fig. 4b, pink versus grey, and Extended Data Fig. 4c, bottom row) and ProteinMPNN sequences generated from the original native backbone (Extended Data Fig. 4b, blue, and Extended Data Fig. 4c, top row). Further, we found in tests on the well-studied colicin–immunity protein system35 that the small changes in protein backbone that partial diffusion can sample are sufficient to mediate specificity changes within protein families (Supplementary Fig. 7). Thus, partial diffusion enables protein backbone resampling and refinement, the extent of which can be tuned by varying the amount of noise added. Furthermore, partial diffusion can considerably increase the designability of input protein models.

As a first experimental test of partial diffusion, we started from our parametrically designed Inpainted binders to GCG (with 231 nM Kd) and NPY (with 3.5 µM affinity; Extended Data Fig. 1a,b). Following partial noising and denoising, we identified diverse designs (Supplementary Fig. 8) that in silico had substantially improved computational metrics compared to the starting design (Supplementary Fig. 9). We used an auxiliary potential during the denoising trajectory8 that minimized the radius of gyration (see Code availability) of the protein–peptide complex to promote additional interaction with the peptide. Initial screening with yeast surface display revealed quite high binding success rates, with 25 out of 96 designs binding GCG, and 20 out of 96 binding NPY at 10 nM peptide concentration. The highest-affinity designs were expressed in E. coli and then purified, and their binding affinities were determined using FP to be 5.6 nM to NPY (Fig. 2b, top) and below the limit of detection in the picomolar range to GCG (Fig. 2b, bottom). The designs were quite specific: the GCG binder bound preferentially to GCG over the closely related SCT, and particularly notably, the NPY binder did not show any cross-reactivity to peptide YY (PYY), a member of the NPY and pancreatic polypeptide family36 with 63% sequence identity to NPY (Fig. 2b).

To gain insight into the structural rearrangements generated by partial diffusion that contribute to the affinity increases, we solved the structures (Extended Data Table 2) of the original Inpainted GCG binder and the partially diffused higher-affinity GCG binder. Both designs were very close to their design models. Subtle structural changes in the protein backbone between the original Inpainted design model and the partially diffused model are nearly perfectly recapitulated in the corresponding crystal structures (Fig. 2c). Alignment of the two crystal structures (Fig. 2d) on the structurally conserved C-terminal residues (16–29) of GCG (Supplementary Fig. 10) showed that in the partially diffused GCG binder a 2.7-Å shift towards the target in the binder backbone enables an isoleucine to fit into a pocket previously occupied by a phenylalanine side chain at position 13 (Fig. 2e, left inset). Similarly, at position 16, a 3.6-Å shift in the backbone allows a tyrosine residue to pack underneath the peptide and form a hydrogen bond to the peptide backbone where previously a serine could not make any contacts (Fig. 2e, right inset). These backbone movements and accompanying sequence changes increase the interaction shape complementarity (0.62 versus 0.67) and contact molecular surface (431 Å2 versus 522 Å2; computed on the crystal structures). We observed similar improvements in estimated binding energy (Rosetta ddG) and contact molecular surface after running partial diffusion starting from the Inpainted designs for GCG and NPY (Supplementary Fig. 11).

De novo binder design using RFdiffusion

Inspired by this success at optimizing binders with RFdiffusion, we next tested its ability to design binders completely de novo through unconditional binder design. We first used the fixed target structure approach of ref. 8, and provided RFdiffusion with the sequence and structures of the two peptides in helical conformations, leaving the topology of the binding protein and the binding mode completely unspecified (Fig. 3a). From this minimal starting information, RFdiffusion generated designs predicted by AF2 to fold and bind to the targets with high in silico success rates. A representative design trajectory is shown for PTH in Supplementary Video 1; starting from a random distribution of residues surrounding the PTH peptide in a helical conformation, in sequential denoising steps the residue distribution shifts to surround the peptide and progressively organizes into a folded structure that cradles almost the entire surface of the peptide.

Fig. 3: De novo peptide binder design with RFdiffusion.
figure 3

a, Schematic showing peptide binder design using RFdiffusion. Starting from a random distribution of residues around the target peptide (XT), successive RFdiffusion denoising steps progressively remove the noise leading at the end of the trajectory to a folded structure, X0, cradling the peptide. At each step t, RFdiffusion predicts the final structure pX0 given the current noise sample Xt, and a step that interpolates in this direction is taken to generate the input for the next denoising step Xt−1. b, Design of picomolar-affinity PTH binder. Top: design model of PTH binder (spectrum, AF2 metrics in Supplementary Table 9). Middle: circular dichroism data show that the binder has helical secondary structure and is stable at 95 °C (inset). Bottom: FP measurements (n = 4) with PTH indicate a subnanomolar binding affinity and no binding to PTHrp indicates high specificity. c, Design of picomolar-affinity Bim binder. Top: design model of Bim binder (spectrum, AF2 metrics in Supplementary Table 9). Middle: circular dichroism data show that the binder has helical secondary structure and is stable at 95 °C (inset). Bottom: FP measurements (n = 4) with Bim indicate a subnanomolar binding affinity. d, Crystal structure of Bim binder (teal and red). Top inset: a cross-interface hydrogen-bond network formed between Asn20 of Bim and Thr73 and Asn77 of the binder. Bottom inset: a kinked helix in the diffused backbone accommodates Arg13 of Bim. e, RFdiffusion with PYY sequence input alone. Left: PYY in complex with its native NPY Y2 receptor37 (PDB ID: 7YON) shows flexibility at its N and C termini (teal). Middle: design model of the binder (spectrum) with PYY target (brown); the peptide is more ordered in both regions (N terminus, teal). Right: FP measurements (n = 4) with PYY indicate a 24.5 nM binding affinity.

Full size image

We obtained synthetic genes encoding 96 designs for each target. Using yeast surface display, we found that 56 of the 96 designs bound to PTH at 10 nM peptide concentration. The highest-affinity design again bound too tightly for accurate Kd estimation; instead FP data provide an approximate upper bound for the Kd of < 500 pM (Fig. 3b, bottom). Binding was also highly specific; no binding was observed to the related PTHrp (Fig. 3b, bottom). For Bim, 25/96 of the designs bound by yeast surface display, and FP on the highest-affinity design indicated a Kd of < 500 pM (Fig. 3c, bottom). Circular dichroism temperature melts indicate that both binders are stable at 95 °C (middle panels of Fig. 3b,c). The completely de novo diffused binders again had considerable structural similarity to our starting groove binding concept (compare the top panels of Fig. 3b,c to the middle panel of Fig. 1b). We solved the X-ray crystal structure (Extended Data Table 2) of the Bim binder, and found that it closely matched the design model (3.0 Å resolution, 0.57 Å RMSD; Fig. 3d). A kinked helix on the binder adjacent to the interface is well recapitulated in the structure, and a cross-interface hydrogen-bond network designed between Thr73 and Asn77 of the binder and Asn20 of Bim forms in the otherwise hydrophobic interface.

We next sought to generalize RFdiffusion to enable binding to flexible targets from a specification of the target sequence alone (as can be achieved with AF2 Hallucination, detailed above). We fine-tuned RFdiffusion by training on two chain systems from the Protein Data Bank (PDB), noising the structure on one and providing only the sequence on the second. We found that the fine-tuned version could readily design folded structures around a variety of peptides given only sequence information. We used this approach to design binders to PYY (Fig. 3e), which in the cryogenic electron microscopy structure with the NPY Y2 receptor is incompletely resolved and adopts a partially helical structure37. Starting from only the amino acid sequence of PYY, RFdiffusion generated solutions with the peptide in a range of conformations. A design with the peptide adopting a different conformation from the experimental structure bound PYY with 24.5 nM affinity (Fig. 3e, right). Note that here we explored using shorter binder chain lengths in these calculations, resulting in smaller designs, which probably accounts for the lower affinity than in the fixed structure case above. Lower affinity binders were also obtained for PTH and GCG using this flexible backbone RFdiffusion approach (Extended Data Fig. 5a,b).

Human versus machine problem solving

The deep learning methods largely converged on the overall solution to the helical peptide-binding design problem—groove-shaped scaffolds with helices lining the binding site—that the human designers chose in the initial Rosetta parametric approaches. The increased affinity of the deep learning designs probably derives at least in part from higher shape complementarity resulting from direct building of the scaffold to match the peptide shape; the average contact molecular surface for the partially diffused GCG binders and NPY increased by 33% and 29% respectively compared to that of the starting models, and the Rosetta ddG improved by 29% and 21% (Supplementary Fig. 11). The ability of RFdiffusion de novo design to rapidly ‘build to fit’ provides a general route to creating high-shape-complementary binders to a wide range of target structures, and as noted above, partial diffusion provides a general route to sampling binders with increased affinity by making small backbone adjustments to enable placement of more space-filling side chains.

Design of protein biosensors

Given our success in generating de novo binders to clinically relevant helical peptides, we next sought to test their use as detection tools for use in diagnostic assays. Compared to immunosensors, de novo protein-based biosensors can offer a more robust platform with high stability and tunability for diagnostics38. To design PTH biosensors, we grafted the 6.1-nM PTH binder into the lucCage system39 (Fig. 4a), screened eight designs for their luminescence response in the presence of PTH, and identified a sensitive lucCagePTH biosensor (LOD = 10 nM) with ≈21-fold luminescence activation in the presence of PTH (Fig. 4b).

Fig. 4: Application of designed binders to sensing and detection.
figure 4

a, The PTH lucCage biosensor. Cage and latch (left, beige), key (right, beige) and the PTH binder (grey) thermodynamically shift from the off to on state in the presence of PTH peptide target (purple). This conformational change brings two luciferase halves (inactive in white, active in blue) close together leading to luminescence. b, Left: titration of PTH results in luminescence increase (n = 3). Middle: response of lucCagePTH biosensor in the linear concentration range, indicating a 10 nM limit of detection (Supplementary Methods). Right: titration curve of 10 nM lucCagePTH + lucKey to various concentrations of PTH (n = 3). c, LC–MS/MS enrichment experiment schematic; the trypsin digestion step was skipped for the GCG binder. d, Left: LC–MS/MS recovery percentages for triplicate measurements of an N-terminal tryptic peptide of PTH. The negative control comprised bovine serum albumin mixed with PTH in a buffer solution. Right: recovery percentage for triplicate measurements of intact GCG peptide normalized to the percentage recovery with a monoclonal antibody (n = 3). Following the first binding and elution experiments, beads were extensively washed and resuspended in PBS–CHAPS 0.1%, and then used in a second pulldown experiment. An unrelated binder attached to the magnetic beads mixed with GCG in buffer was used as a negative control. a,c, Created with BioRender.com.

Full size image

Enrichment for LC–MS/MS detection

We explored the use of our picomolar-affinity RFdiffusion-generated binders to PTH and GCG as capture reagents in immunoaffinity enrichment coupled with liquid chromatography–tandem mass spectrometry (LC–MS/MS), a powerful platform for detecting low-abundance protein biomarkers in human serum40. We prepared PTH- and GCG-binder-conjugated beads as described in the Methods. PTH enrichment was quantified on the basis of the analysis of the amino-terminal peptide of a tryptic digestion of PTH in human plasma41 (Methods and Extended Data Fig. 6a). We found that the designed binder enabled capture of PTH from buffer and human plasma supplemented with PTH (the endogenous levels are too low for reliable detection) with recoveries of 53% and 43%, respectively (Fig. 4d, left). For GCG, enrichment was quantified for recovery of peptide in buffer solution (see Methods and Extended Data Fig. 6b) because recovery was low in extract (further increases in specificity of this design will probably be necessary for actual applications). The GCG binder beads had comparable peptide capture efficiency to that of monoclonal GCG antibody beads, with 91.1% recovery when normalized to the antibody’s 100% recovery rate in a spiked buffer (Fig. 4d, right). In contrast to the antibody-coupled beads, which lost almost all GCG-binding activity after the first use (Fig. 4d, right), the GCG-binder-conjugated beads retained almost full binding activity in a second capture experiment (Fig. 4d, right). This greater robustness to washing and repeated use probably reflects the exceptional stability of the designed binders (middle panels of Fig. 3b,c and Extended Data Fig. 3e), which could substantially lower cost (as they are no longer single use) and extend shelf life compared to using antibodies.

Discussion

Antibodies have served as the industry standard for affinity reagents for many years, but their use is often hampered by variable specificity and stability10,11. For binding helical peptides, the computationally designed helical scaffolds described in this paper have a number of structural and biochemical advantages. First, the extensive burial of the full length of an extended helix is difficult to accomplish with antibody loops15, but very natural with matching extended α-helices in groove-shaped scaffolds. Second, designed scaffolds are more amenable to incorporation into sensors as illustrated by the lucCagePTH sensor. Third, they are more stable than antibodies, can be produced much less expensively, and can be easily incorporated into affinity matrices for enrichment of peptide hormones from human serum (the striking difference in the robustness of antibody-conjugated versus binder-conjugated beads to repeated use (Fig. 4d, right) highlights the differences in stability of the two modalities). Fourth, computational design avoids the need to immunize animals, which often mount weak responses to highly conserved bioactive molecules42. MS-based detection of peptides following enrichment using designed binders could provide a general route forwards for serological detection of a wide range of disease-associated peptide biomarkers.

Our results highlight the emergence of powerful new deep learning methods for protein design. The RFjoint and RFdiffusion methods were both able to improve on initial Rosetta designs, and the Hallucination approach generated high-affinity binders without requiring pre-specification of the bound structures. Moreover, the RFdiffusion method rapidly generated very tight (picomolar Kd values) affinity and specific binders to several helical peptides. RFdiffusion was previously shown to be able to design binders to folded targets8; here we demonstrate further that it can be used to improve starting designs by partial noising and denoising, and can generate binders to peptides starting from no information other than the target sequence. To our knowledge, the Bim- and PTH-binding proteins diffused starting from random noise are the highest-affinity binders to any target (protein, peptide or small molecule) achieved directly by computational design with no experimental optimization. We expect both the RFdiffusion de novo peptide binder design capability and the ability to resample around initial designs (before or after experimental characterization) to be broadly applicable.

Methods

Computational methods

Parametric design of groove-shaped scaffold library and use for binder design

The parametric groove-shaped scaffold library was sampled using a random sampling approach, for which key parameters22 were selected randomly from specific distributions. An even distribution of bundle ‘lengths’ was sampled, for which each parametric helix was 15–19 residues long. A supercoiling value was randomly selected from a biased distribution favouring more supercoiled scaffolds, given that these scaffolds were more likely to fail in the subsequent looping step (Fig. 1b and Supplementary Fig. 1). This biased sampling strategy was chosen to achieve a more uniform distribution of supercoiling within the final scaffold library, with sufficient numbers of highly supercoiled bundles. An average helix neighbour distance value was randomly selected from a Gaussian distribution informed by native protein helical bundle geometries (Fig. 1b and Supplementary Fig. 1). The distance of each helix from its neighbours was independently randomly selected from a much tighter Gaussian distribution centred at the preselected average helix neighbour distance value, to provide some noise within a given scaffold to helix distances and allow for heterogeneous amino acid selections (Supplementary Fig. 2). Values for helix phase and z displacement were randomly sampled for each helix. The ‘groove’ consisting of three helices was first sampled as a helical bundle using the Crick parameterization of α-helical coiled coils, around an imaginary central helix where the target was to later be docked. Next, the two buttressing helices were sampled with the same parameterization, but moved radially outwards with randomly sampled helix neighbour distances as well as an additional randomly sampled tilt. This process was used to sample a set of 200,000 arrangements of 5 helices. Next, the Rosetta ConnectChainsMover51 was used to loop this set into approximately 135,000 successful scaffold backbones. These backbones were designed and filtered using Rosetta52 (including flexible backbone design) to yield a final library of 18,000 scaffolds. Backbones were filtered on metrics including buried nonpolar surface area per residue, Rosetta score per residue, percentage alanine, exposed hydrophobics per residue, and Rosetta ‘holes’53. This library was used to design binders to different helical peptide targets using an adapted version of the miniprotein binder design computational pipeline used in ref. 7, in which only the binder interface was designed and the target was restricted to only rotamer repacking.

RFjoint Inpainting

To sample around an initial putative binder, and to extend the binding interface to make additional contacts with the bound peptide, the RFjoint Inpainting network was used23, in conjunction with ProteinMPNN24. Rosetta-designed binders to PTH, GCG and NPY were used as input to RFjoint. RFjoint is deterministic, and hence, to generate diversity, additional length was added (randomly and independently sampled) at the loop junctions between the binder helices. Additionally, one whole helix was completely rebuilt by RFjoint, to further permit diversification. RFjoint designs were subsequently sequence redesigned with ProteinMPNN, validated and filtered in silico by AF2 with initial guess6,28, and subsequently tested experimentally.

Sequence threading to generate peptide binders

We started from a library of several thousand all-helical scaffolds bound to designed single helices. We then threaded sequences of peptides of interest onto the bound single helix and filtered to obtain threaded conformations that maximized the number of target sequence positions that formed hydrophobic interactions at the interface to the binder scaffold17,26. The resulting binders were then redesigned in the presence of the threaded target sequence with ProteinMPNN24 (forbidding cysteine) and the complex was predicted with AF2 with initial guess6,28. Another round of ProteinMPNN and AF2 + initial guess was carried out on the AF2 models that passed gate filters. Both rounds had gate filters of interface pAE < 10, mean pLDDT > 92, predicted template modelling (pTM) score > 0.8 and RMSD to input backbone <1.75 Å. AF2 models from both rounds that passed gate filters were further filtered on AF2 metrics and filtered on Rosetta metrics to select sequences to order. Sequences were filtered against membrane insertion potential54, contact_molecular_surface, ddG7, interface pAE and monomer pAE6.

AF2 Hallucination for flexible peptide binder design

Code for running Hallucination with AF2 was modified from ref. 31, with custom losses developed to promote binding of the Hallucinated protein to the input peptide sequence. AF2 model_4_ptm was used for all experiments.

Initial sequence sampling

In line with ref. 31, the initial binder sequence was sampled randomly, with amino acid probabilities corresponding to background amino acid frequencies in BLOSUM6255. The target sequence (but no template structure) is also provided, separated by a chain break (+32 residue positional index offset). Residues were then mutated, with probabilities related to their background frequency in BLOSUM62. The number of amino acid changes proposed at each step is decayed throughout the trajectory (1,250 ×3 steps, 2,500 ×2 steps, 1,250 ×1 step). Multiple simultaneous amino acid changes initially helps speed up Hallucination, and a lower rate of changes later on allows more gradual refinement. To further speed up convergence, alterations were selectively made to residues with the lowest 50% of AF2 pLDDTs.

Losses used for Hallucination

Losses used for Hallucination were as follows.

  • pLDDT of the bound state: average pLDDT of the binder–peptide complex.

  • pTM of the bound state: the pTM score of the binder–peptide complex.

  • Radius of gyration: the radius of gyration was calculated as the mean squared distance of residues from the centre of mass of the protein. To approximately standardize the scaling with length of the protein, this was empirically normalized by dividing the radius of gyration by the radius of a sphere of volume related to the length of the Hallucinated protein.

  • Contact probability: calculated as total probability that a residue in the target is in contact (closer than 8 Å) with the target peptide (the summed probability over the sub-8-Å bins of the distogram output from AF2). This was averaged across all target residues.

  • Interface pAE: the mean pAE between the binder and peptide chains.

For all examples shown in this work, the losses were weighted with relative weights of 1:1:0.1:3:5.

Simulated annealing

To optimize the designed binder, Monte Carlo simulated annealing was carried out, with a starting temperature of 0.01, and the half-life of the exponential decay set to 500 steps. Alterations were accepted or rejected using the Metropolis criterion. A total of 5,000 steps were carried out during design.

ProteinMPNN

Previous work has demonstrated that AF2 Hallucination yields adversarial sequences that do not work experimentally31. However, designs can be rescued with ProteinMPNN redesign of the sequences. Sixty-four sequences were designed per backbone, and were subsequently filtered on the basis of AF2 pLDDT, pTM, RMSD to the design model, RMSD of the monomer to the binder model (without the peptide), and Rosetta ddG. The precise values used for filtering were chosen to reduce the set down to 46 designs.

Partial diffusion to optimize binders

RFdiffusion was modified to allow the input structure to be noised only up to a user-specified time step instead of completing the full noising schedule. The starting point of the denoising trajectory is therefore not a random distribution. Rather, it contains information about the input distribution resulting in denoised structures that are structurally similar to the input (Fig. 2a). The AF2 models of the highest-affinity designs from Inpainting for GCG and NPY were used as inputs to partial diffusion. The models were subjected to 40 noising time steps out of a total of 200 time steps in the noising schedule, and subsequently denoised. An auxiliary potential minimizing the radius of gyration of the binder–peptide complex was used (described below). Approximately 2,000 partially diffused designs were generated for each target. The backbones in the resulting library were sequence designed using ProteinMPNN (and ProteinMPNN after Rosetta FastRelax), followed by AF2 + initial guess6. The resulting libraries were filtered on AF2 pAE, pLDDT, RMSD to the design model, RMSD of the monomer to the binder model (without the peptide) and Rosetta ddG. The precise values used for filtering were chosen to reduce the set down to 96 designs for each target.

De novo peptide binder design using RFdiffusion

The AF2 model of the PTH peptide in the highest-affinity binder from Inpainting was used as input to RFdiffusion. For Bim, there was no previously designed binder and therefore the crystal structure of Bim56 (PDB: 6X8O) was used as input. An auxiliary potential minimizing the radius of gyration of the binder–peptide complex was used during denoising (described below). Approximately 2,000 diffused designs were generated for each target. The backbones in the resulting library were sequence designed using ProteinMPNN (and ProteinMPNN following FastRelax), followed by AF2 + initial guess6. The resulting libraries were filtered on AF2 PAE, pLDDT, RMSD to the design model, RMSD of the monomer to the binder model (without the peptide) and Rosetta ddG. The precise values used for filtering were chosen to reduce the set down to 96 designs for each target.

Radius of gyration potential

RFdiffusion enables the use of external guiding potentials during inference, which help in the design of proteins with a certain desired property. The utility of these guiding potentials in designing symmetric oligomers and enzymes, as well as a description of how they are incorporated into the sampling procedure, is reported in ref. 8. In this work, we take advantage of guiding potentials to minimize the radius of gyration of the binder–peptide complex. The radius of gyration is calculated as the root mean square of the distance of all the Cα atoms from the centroid. It is more important to apply the potential at the initial denoising steps, and less so towards the end when the quaternary structure is largely fixed. Therefore, the scaling factor with which the gradients are multiplied has a cubic decay over the course of the denoising trajectory.

Training RFdiffusion for designing binders to targets from sequence alone

A modified version of RFdiffusion was trained to permit the design of protein binders to targets, for which only the sequence of the target was specified. The training strategy largely followed the training strategy used for the original RFdiffusion model, with some modifications. A summary is provided below.

Overview of ‘base’ RFdiffusion training

RFdiffusion8 is a denoising diffusion probabilistic model fine-tuned from a pretrained structure prediction model; RoseTTAFold57,58. RFdiffusion is trained with a forward noising process that iteratively, over 200 time steps, noises residue translations and orientations to distributions that are indistinguishable from random distributions (three-dimensional Gaussian distribution and a uniform distribution on SO(3), respectively). RFdiffusion is then trained to reverse this corruption process, predicting the ground truth (X0) at each time step of prediction. Mean-squared-error losses are used to minimize the error between the forward and reverse processes. Full training details are extensively described in ref. 8.

Modifications to RFdiffusion for binder design to sequence inputs alone

RFdiffusion was trained on both monomers (<384 amino acids) and heterocomplexes (one chain, denoted the ‘binder chain’, <250 amino acids) from the PDB. Coordinates were scaled by a factor of four, in line with the original RFdiffusion model. In 20% of cases, no sequence or structure was provided to the model (for unconditional generation). In the other 80% of cases, 20–100% of the protein was noised. In contrast to RFdiffusion, however, the structure of up to 50% of the protein (monomer or ‘target chain’) was noised (diffused), while providing the sequence of those residues. Thus, RFdiffusion learns to condition its predictions on the sequence (without structure) of part of a protein (the monomer) or of a target to bind to. This version of RFdiffusion was trained for seven epochs.

Computational filtering

Precise metrics cutoffs changed for each design campaign to get to an orderable set, but largely focused on pAE (<10), pLDDT (>80) and Rosetta ddG (<−40)6.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.