A Peptide With No Shape Was Just Gripped Tighter Than the “Easy” Case
Dynorphin A is seventeen residues long, and in solution it does not have a structure. Not “a flexible structure” or “a structure that breathes” — nuclear magnetic resonance shows it sampling a shifting population of backbone conformations with no single dominant fold, the textbook definition of an intrinsically disordered peptide [3]. It is also the endogenous agonist of the kappa opioid receptor (KOR), a signaling axis with real, stated pharmaceutical interest for pain relief without the addictive liability of mu-opioid drugs, and by its own discoverers’ account: “To date, no antibodies, peptides, or small molecules have been developed to inhibit dynorphin A; existing ligands instead modulate KOR signaling by engaging the deep binding pocket of the receptor” [3]. Nobody had built a molecule that grips the free peptide itself.
In 2025, David Baker’s lab published two separate papers, months apart, both aimed at exactly this class of problem — a target with no single shape to complement. One method reaches a set of real disordered proteins at 3-to-100-nanomolar affinity [1]. The other, applied to dynorphin A specifically, produced a design later measured below 60 picomolar by fluorescence polarization [3] — tighter than the same broader RFdiffusion lineage’s own headline “easy,” fully-folded comparison cases, which bottom out around 500 picomolar [5]. That is not a small distinction. It inverts the assumption this entire subfield started from: that a target without a fixed structure to complement is the harder design problem, not the easier one. The two 2025 papers, read against each other and against the folded-target baseline that predates both, are the first place that assumption gets a number attached to it rather than an intuition.
Two Papers, One Problem, Two Different Machines
Both papers start from RFdiffusion, the diffusion-based generative model Watson and colleagues built by fine-tuning RoseTTAFold’s structure-prediction network on a protein-backbone denoising task, originally demonstrated on folded monomers, symmetric assemblies, and binders to ordered targets [6]. Neither paper treats that architecture as something to modify at the algorithmic level — both cite it exactly once, as the tool, and then spend their own pages on what they built with it. That is the right division of labor for this article too: RFdiffusion’s own training and sampling procedure is TI’s territory, not this one’s. What each paper does with it, on a target that has no fixed shape, is the object here.
“Diffusing protein binders to intrinsically disordered proteins” takes the direct route: freely sample the conformation of both the target and the binder at once during the diffusion process itself, rather than fixing either one in advance [1]. For longer targets — amylin’s 37 residues, C-peptide, the 39-amino-acid viral transactivator VP48 — this means feeding only the target’s sequence into the model and letting it discover, jointly with a candidate binder scaffold, which of the many conformations that sequence can adopt happens to pair with a bindable protein surface. For shorter disordered windows, the same paper adds the option to specify a target secondary structure directly — helix, strand, or loop — narrowing the search when a plausible bound geometry can be guessed in advance.
“Design of intrinsically disordered region binding proteins,” published in Science the same month, takes a structurally different route the authors name “logos” [3]. Instead of sampling target and binder conformation jointly from noise every time, the lab first built a fixed library of 1,000 template structures — extended protein scaffolds, each with a defined series of binding pockets shaped to wrap around a peptide backbone in an extended, non-folded conformation, assembled in a three-step pipeline of Rosetta-based scaffold generation, diffusion-based pocket specialization, and RFdiffusion-linked pocket assembly. Given a new disordered target, the method threads its sequence onto every template in that library to find compatible segment-template pairs, then runs a further, cheaper round of diffusion refinement only on the best matches. David Baker put the resulting division of labor plainly: “The RFdiffusion-based method excels at designing binders to targets with some helical and strand secondary structure, while the logos method works best for targets lacking regular secondary structure” [8]. These are not two attempts at the same recipe. They are two different bets on where a disordered target’s usable information actually lives — in its plausible secondary structure, or in the raw geometry of its side chains regardless of backbone shape — tested against real, synthesized, measured molecules within the same calendar year, by overlapping author lists, from the same institute.
Before pointing the library at any real biological target, the Science paper ran a sanity check with no biology attached to it at all: binders designed against six arbitrary English words and names, encoded as peptide sequences with no evolutionary history and no natural binding partner. Eight designs were tested per target on average; the best binders for two of the six words reached single-digit nanomolar affinity (Kd = 9 nM), three more landed in the double digits (35, 37, and 90 nM), and the last came in at 180 nM [3]. A method that works on a sequence with no biological reason to be bindable at all is a method being tested against the null hypothesis that its later, real-target successes are secretly exploiting some evolved, biology-specific recognition feature. That null hypothesis did not survive the check.
The Free-Diffusion Route Tops Out at 100 Nanomolar, With the Receipts to Show Why
Take the Nature paper’s numbers first, because they are the more intuitive result — the one that matches the assumption this article opened by stating. Screening 96 initial designs against amylin by bio-layer interferometry returned affinities from 100 to 454 nanomolar; a second round of 174 refined designs pushed 107 of them (61 percent) into confirmed binding, and the four best-characterized final binders span 3.8 to 100 nanomolar depending on which conformation of amylin each one targets — Amylin-68nαβ at 3.8 nM, Amylin-36αβ at 10 nM, Amylin-75αα at 15 nM, Amylin-22αβL at 100 nM [1]. C-peptide fared worse by hit rate: of the initial 96 designs tested, only one showed any binding at all, and weakly; a further round of the same two-sided partial diffusion refinement applied to that single hit produced six designs binding better than the 100-nanomolar threshold, the best, CP-35, at 28 nM. VP48 was the hardest of the three: only 2 of 95 initial designs bound, at roughly 750 nanomolar, and it took a further optimization round to reach the reported 39 nM. The paper’s own three IDR cases — G3BP1’s RNA-binding domain, the IL-2 receptor γ-chain, and the prion protein’s amyloid core — followed the same shape: G3bp1-11 reached 11 nM after a jump from 5-of-78 initial hits to 40-of-95 (42 percent) in a refined round; IL2RG-30 was optimized from an initial 493 nM down to 97 nM; PRI28, built by specifying the prion fragment’s secondary structure as a beta strand rather than sampling it freely, reached 14 nM out of a first screen of 48 designs.
None of this is a soft result. Two of the nine reported complexes were solved by X-ray crystallography, and the structures match the design models closely: Amylin-22αβL bound to amylin at 1.8 Å resolution, with an interface Cα RMSD of 1.33 Å and sidechain RMSD of 1.87 Å against the computational model; G3bp1-11 bound to the G3BP1 domain at 2.4 Å, with a whole-complex Cα RMSD of 0.8 Å [1]. The functional tests hold up too. At a 1-to-4 binder-to-amylin molar ratio (10 μM binder, 40 μM amylin), every one of the four amylin binders completely inhibited new fibril formation by thioflavin-T fluorescence, and the same binders dissolved pre-formed fibrils in a concentration-dependent manner — a designed protein interrupting, in a test tube, the same aggregation process that deposits in a type 2 diabetic pancreas. As a mass-spectrometry capture reagent, the amylin binder recovered 62.2 percent of spiked amylin from buffer and 53.5 percent from human plasma, where endogenous concentrations are otherwise too low for reliable detection [1]. This is a real, working, structurally confirmed method. Its affinity ceiling, on the evidence assembled here, sits at 100 nanomolar for the hardest case and rarely below 3.8 nanomolar for the easiest.
The Logos Method Reached Sixty Picomolar Against a Target Nothing Else Had Ever Bound
The Science paper’s headline numbers, in its final published form, cover 39 diverse disordered targets and report 34 successes at affinities from roughly 100 picomolar to 100 nanomolar, testing about 22 designs per target on average [2]. That range alone already overlaps and exceeds the Nature paper’s ceiling. Dynorphin A is where the overlap turns into something sharper.
An initial threaded design, DYNA_1b1, bound dynorphin A at roughly 1 nanomolar. The authors then applied a further round of RFdiffusion-based refinement on top of that hit — the same kind of local, structure-aware refinement step used elsewhere in the paper on synthetic targets. Forty-five of 48 refined designs showed strong binding in a bio-layer interferometry screen run at a 5-nanomolar peptide concentration; six of those forty-five had dissociation constants at or below 100 picomolar by the same assay. Two of those six were measured again by the more sensitive method of fluorescence polarization: DYNA_2b1 came back below 60 picomolar, and DYNA_2b2 below 200 picomolar [3]. Put next to the RFdiffusion lineage’s own previously reported “easy” cases — parathyroid hormone, the pro-apoptotic peptide Bim, and glucagon, all folded-upon-binding helical peptides whose designed binders reached affinities below the roughly 500-picomolar limit of detection for the fluorescence-polarization assay used to measure them [5] — a binder against a peptide with no fixed shape at all now sits at the same order of magnitude, or tighter, than binders against peptides that fold into a single clean helix the moment they’re asked to.
A co-crystal structure closes the gap between “measured” and “understood.” DYNA_1b7, a related 7-nanomolar dynorphin A binder from the same optimization series, was solved in complex with dynorphin A residues 1 through 17 at 3.15 Å resolution — deposited to the Protein Data Bank as entry 9CCE on June 21, 2024, publicly released August 13, 2025 [4]. Every hydrogen bond the design model predicted from the binder’s Asn19, Asn69, and Asn70 to the peptide backbone was present in the crystal exactly as designed; the corresponding peptide region, from Leu5 to Arg9, aligned to the design model with a Cα RMSD of 0.6 Å [3]. To gain specificity against the closely related dynorphin B and other neuropeptides, the design deliberately excluded the shared N-terminal Tyr-Gly-Gly-Phe motif those peptides all share — a targeting choice, not a limitation the structure exposed after the fact. Alanine-scanning across nine positions on the peptide, tested against the earlier DYNA_1b1 design in a split-luciferase binding assay, found that every single substitution considerably weakened binding — the designed contacts are load-bearing, not a lucky, non-specific stick [3]. And the tightest design does something a structure alone can’t show: in a cAMP assay run in cells stably expressing the human kappa opioid receptor, DYNA_2b2 blocked dynorphin A-dependent KOR signaling with an IC50 of 50 nanomolar — a number higher than its own Kd, exactly as expected, because it is now competing directly against the receptor itself for the same disordered peptide [3].
What the NMR Spectrum Actually Proves, and What It Doesn’t
The dossier question this article was built to answer is a mechanistic one: when a designed protein binds a target with no fixed structure, does it capture one specific conformation out of the target’s accessible ensemble, or does it make an averaged, fuzzy contact across many states at once? For dynorphin A, the paper does not leave this to inference. Isotope-labeled dynorphin A was run through NMR in three conditions: free in solution, bound to the 7-nanomolar DYNA_1b7, and bound to the sub-200-picomolar DYNA_2b2. Free dynorphin A came back disordered, exactly as expected. Bound to either design, it came back ordered — in both cases in an extended conformation consistent with the respective design model — with the degree of ordering measurably greater for the tighter DYNA_2b2 complex, particularly around the C-terminal region and a Trp14-Asn137 bidentate hydrogen-bond interaction, consistent with that design’s more extensive backbone hydrogen-bonding network [3]. That is conformational capture stated as a direct experimental result, not an interpretive gloss on a computed model: one peptide, two measured states, and a designed binding partner that visibly selects and stabilizes one specific conformer out of what was, a moment before binding, a shifting population.
The second half of that result is where the essay’s real discriminator sits. The extended conformation the designed binders capture is not the conformation dynorphin A adopts on its own natural target. Previously solved cryo-EM and NMR structures of dynorphin bound to the native kappa opioid receptor (PDB entries 2n2f and 8f7w) show the peptide in a compact, partial-helix fold — a categorically different geometry from the fully extended chain the crystal structure of DYNA_1b7 and the bound-state NMR of DYNA_2b2 both confirm [3]. This is not a design flaw. The two binding events are solving different problems: the receptor recognizes dynorphin A in whatever conformation triggers its own downstream signaling machinery, while the designed binder was never asked to reproduce that state — only to find some state, out of the peptide’s full accessible range, that a protein pocket could grip tightly and specifically. The two papers’ own amylin structures point at the same pattern from the other direction: Amylin-22αβL captures amylin in one specific α-β-loop arrangement among the four different conformations the paper’s four named amylin binders separately captured, each one crystallographically or computationally distinct from the others [1]. Across every structurally solved case in both 2025 papers, the pattern is consistent: one binder, one captured conformer, confirmed at atomic or near-atomic resolution. No solved structure in either paper shows an averaged, multi-state interface. The counter-case this entry’s own architecture flagged in advance — that some binders might turn out to contact a genuine ensemble rather than one selected state — does not survive contact with the structural data currently available. It may still turn out to describe some future case; it does not describe any of the nine structurally or NMR-characterized cases reported so far.
The Recognition-Geometry Ratio Is Not One Number, and That’s the Actual Finding
Put a number on the comparison this article has been building toward. Define a recognition-geometry ratio R as the best reported dissociation constant against a disordered target, divided by the best reported dissociation constant against a comparably screened, RFdiffusion-lineage folded or order-upon-binding target — using the roughly 500-picomolar detection floor reported for parathyroid hormone, Bim, and glucagon as that folded-target baseline [5].
For the free-conformation diffusion method applied to amylin — a 37-residue peptide whose four best characterized binders span its full reported range — R runs from about 7.6 (3.8 nM ÷ 0.5 nM, the easiest amylin case) up to about 200 (100 nM ÷ 0.5 nM, the hardest, crystallographically solved amylin case). By this measure, genuinely free-sampled recognition of a longer, conformationally rich disordered protein costs a real, substantial affinity penalty relative to the folded-target baseline — the intuitive result, quantified.
For the logos method applied to dynorphin A — a 17-residue peptide targeted through a pre-built extended-scaffold library rather than free joint sampling — R runs to roughly 0.12 or tighter (60 pM ÷ 500 pM). By the same measure, a genuinely disordered target was bound more tightly than the “easy” folded comparison case, not less. Baker’s own account of why matches the mechanism the NMR data already established: disorder makes binding “almost become easier” than for ordered proteins, because “even though the protein is unfolded on its own, it will adopt a specific structure when it interacts with the designed binder,” and “the binder essentially forces the peptide into a conformation which is well suited to be bound” [7]. Kejia Wu’s account of the same result, in the Science paper’s own discussion, states the same claim as a general design principle rather than a lucky anecdote: “We demonstrate that the conformational heterogeneity of IDPs and proteins can be exploited to make the binder design problem easier than for traditional stable folded targets. For each target, we sampled a wide variety of conformations and identified those compatible with high affinity binding — this contrasts with folded targets, whose fixed structures may admit few optimal binding solutions” [3].
There is a second, independent test of whether that ceiling tracks disorder itself rather than something more specific, and the Science paper ran it directly. Polar residues are traditionally a harder packing problem for protein design tools than hydrophobic ones, so a target’s polarity is a reasonable alternative candidate for what actually sets the difficulty. Twenty of the 39 targets were more than 50 percent polar, six more than 75 percent, and the paper found that reaching those more polar targets required proportionally more hydrogen bonding to the target’s own side chains — 4.3 such bonds per ten residues of target for the less-polar majority, 12.3 for the most-polar tier, roughly a three-fold increase [3]. But that extra structural cost did not show up in the affinities actually achieved: “There was little correlation between binding affinity and the polarity or intrinsic secondary structure of the target; Kds less than 10 nM were achieved for targets with a wide range of polarities and secondary structure propensities… indicating the generality of the method” [3]. That rules out the next most obvious single-axis explanation. The ceiling this article has been computing does not track how disordered a target is, and it does not track how polar it is either — only, on the evidence assembled here, the size and conformational breadth of the specific window a method has to search.
A recognition-geometry ratio that runs from 0.12 to 200 depending entirely on which 2025 paper and which specific target you pick is not a measurement error to be averaged away. It is the actual finding. “Disordered” is not one design difficulty; it is a spectrum of residual conformational freedom at the moment of binding, and the two methods sit at different points on that spectrum by construction. The logos method’s pre-built library of extended, side-chain-facing scaffolds effectively pre-solves a large fraction of a short window’s conformational search before the target sequence is even known, leaving the algorithm to search mainly over a well-behaved, extended backbone family. The free-diffusion method, by contrast, samples a longer target’s full, unconstrained conformational space jointly with the binder every time — a search over a genuinely larger space, for a peptide with more internal degrees of freedom to begin with. The ratio tracks that difference in what each method actually has to search over, not “disorder” as a yes-or-no property of the target.
What Thirty-Nine Targets Don’t Yet Prove
The honest limits of this result sit exactly where the method’s own design choices sit. Every one of the Science paper’s 39 attempted targets is a short window — eight to forty amino acids — deliberately selected from a longer disordered region by first discarding low-complexity segments and sequences with multiple close matches elsewhere in the proteome, specifically to avoid designs that would cross-react [3]. That is a sound engineering choice, and the paper is explicit that it is a choice: FAM21, a roughly 1,000-residue disordered protein tail that is part of the pentameric WASH actin-nucleation complex, is targeted through one selected 27-amino-acid window, not end to end — and a binder to that one window, FAM21_1b1, retrieved the entire five-subunit WASH complex from cell lysate by immunoprecipitation, which is a genuinely useful result but not evidence that the remaining roughly 970 residues of FAM21’s disordered tail are equally targetable. The fraction of binders showing any measurable signal at 500 nanomolar, across all 39 targets, ranged from 2 to 67 percent depending on the target [3] — a wide spread that the paper reports plainly rather than folding into one averaged success rate, and a real signal that not every disordered window is equally tractable even within this method’s own reach. The specificity testing has the same honest scope. Twenty targets — sixteen native disordered proteins plus four synthetic controls — were checked all-by-all in a bio-layer interferometry panel and found tight only to their intended partner up to 1 micromolar cross-reactivity; six binder-target pairs were separately confirmed orthogonal in a live-cell mitochondrial colocalization assay [3]. That is a real specificity result, and it is a twenty-by-twenty and six-pair result, not a whole-proteome one. Every functional application reported so far — the WASH-complex pulldown, a PER2 circadian-protein pulldown, a Mesothelin cell-surface localization test on two cell lines, a 90-percent-recovery diagnostic capture of a mutant cystinosin peptide from buffer and blood [3], and dynorphin A’s own KOR-signaling inhibition — is a cell-lysate, heterologous-cell-line, or in vitro biochemical result. None of it has yet been tested in an intact animal, and the paper does not claim otherwise. The amylin fibril-dissolution result sits at the same evidentiary stage: a test-tube demonstration against a purified peptide, not yet a claim about reversing amyloid deposition in a living pancreas.
What Changes If the General Case Holds
If the pattern established here — that a target’s own residual conformational freedom, not its disorder as a category, sets the achievable affinity ceiling — continues to hold as the method is pointed at more targets, the practical consequence is specific rather than sweeping. It does not mean every intrinsically disordered protein in the proteome just became druggable. It means a defined, repeatable recipe now exists for a specific, checkable subproblem: given a disordered region, find a short, non-degenerate window within it, and there is now a real chance — not a guarantee, on the honest 2-to-67-percent hit-rate spread already reported — of reaching a binder in the nanomolar-to-picomolar range against it, verifiable the same way dynorphin A’s was verified, by NMR-confirmed disorder-to-order transition and, where a crystal will grow, by X-ray structure. The named targets already attempted read like a cross-section of exactly the disease categories where a fixed-structure assumption has stalled drug and diagnostic development for decades: a fusion-oncoprotein fragment from Ewing sarcoma’s EWS/FLI, a malaria circumsporozoite-protein fragment, a cell-surface cancer marker, a lysosomal-storage-disease diagnostic peptide, a circadian-clock scaffolding protein, and a chronic-pain neuropeptide that, by its own discoverers’ account, had never had an antibody, peptide, or small molecule built against it at all [3]. Whether that specific, bounded claim generalizes to the much larger set of disordered regions nobody has tried yet is not something either 2025 paper can answer in advance — it is something the next round of chosen windows, screened the same way, at the same honest hit rate, will have to answer instead.
