A pocket-centric framework for selective targeting of amyloid fibril polymorphs

  1. Gaëtan Ossard
  2. Constantin Bogdan Ciambur
  3. Zachariah P. Schuurs
  4. Ronald Melki
  5. Olivier Sperandio
  6. Eugénie Romero

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

The rapid expansion of high-resolution cryo-EM structures of amyloid fibrils has transformed our understanding of fibril polymorphism, yet it has not been matched by comparable progress in the rational development of protein-selective or polymorph-specific amyloid ligands. One possible explanation is that ligand selectivity is governed not only by global fibril folds, but also by local surface pockets accessible to small molecules. Here, we present a systematic analysis of 400 cryo-EM structures of amyloid-β, tau, and α-synuclein fibrils. Using a unified pocket similarity index and minimum spanning tree representations, we construct global and protein-specific graph representations of the amyloid binding pocket space, and examine how surface cavities are distributed across proteins, polymorphs, and structural contexts. We find that many detectable pockets are shared across multiple fibrillar folds and, in several cases, across distinct amyloid-forming proteins, suggesting that pocket-level convergence may contribute to the limited selectivity of amyloid-directed ligands. Conversely, only a restricted subset of pockets occupies isolated regions of pocket similarity space, defining rare structural opportunities for protein-selective or polymorph-restricted targeting. Analysis of structures of extracted fibrils further shows that disease-derived fibril pockets do not form a completely isolated pocketome subset, but can resemble pockets observed in selected in vitro polymorphs. Together, these results reframe amyloid ligand development as a problem of pocket-level discriminability within a constrained fibril landscape, and provide a structural framework for identifying promising binding sites while avoiding intrinsically non-discriminatory pockets.

Significance Statement

Despite major advances in cryo-EM structure determination of amyloid fibrils, the development of selective ligands for amyloid assemblies remains challenging. By systematically comparing surface binding pockets across 400 amyloid-β, tau, and α-synuclein fibrillar structures, we show that many ligand-accessible cavities exhibit similar geometric and physicochemical properties across fibrillar polymorphs made of distinct proteins. This pocket-level convergence provides a structural basis for understanding why many amyloid ligands exhibit broad binding profiles, while also identifying rare pockets that are sufficiently isolated to support more selective targeting strategies. Our work establishes a pocket-centric framework for interpreting amyloid ligand selectivity and for prioritizing fibril binding sites in imaging and therapeutic ligand development.

Article activity feed

  1. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Response to the Reviewers

    Manuscript number: RC-2026-03632

    Corresponding authors: Olivier Sperandio and Eugénie Romero

    We thank the Editor and the three reviewers for their careful and constructive assessment of our manuscript. We are grateful for the positive feedback on the concept of a pocketome framework for comparing local binding environments across amyloid fibrils. The comments have also helped us clarify an important distinction that was not sufficiently explicit in the original manuscript: our analysis characterizes recurrent structural and physicochemical pocket environments, but does not by itself predict ligand affinity or selectivity. In the revised manuscript, we therefore narrow the interpretation of the pocketome and strengthen the study through experimental ligand-site validation, repeatability and calibration analyses, sensitivity analyses, and an external validation on TDP-43, TMEM106B and transthyretin. We also clarify that the reference pocketome is constructed from amyloid-β, tau and α-synuclein and represents the sampled ordered-core surface environments rather than an exhaustive catalogue of all amyloid ligand-binding modes.

    Reviewer #1

    1. __ Although this is a computational study, there is a lack of validation. This paper is mainly focused on alpha-synuclein however it would be useful if the authors could validate on tau where ligands such as APN-1607 and MK-6240 have been found to bind to AD PHFs and SFs by cryo-EM. They could also test AV1451 which binds to CTE filaments by cryo-EM. __We thank the reviewer for this helpful and important comment. We have assembled 38 ligand-bound amyloid fibril structures, including tau structures containing APN-1607, MK-6240 and flortaucipir/AV-1451. For each structure, the ligand is removed before pocket detection and the ligand-bound structure is used only afterwards to identify whether a detected pocket corresponds to the experimentally occupied site. We will report site recovery, spatial overlap and filter survival, including which filter removes experimentally occupied sites when they are not retained. This provides a direct validation of the pocket-detection procedure without retuning the detector to individual ligands.

    2. __ Robustness of the pocket detection pipeline. If you change the rotamer of a side chain where the ligand has been observed to bind or the charge state, does the pocket remain stable? I suppose one could test this on the high redundancy of the same in vitro structures that reappear in the PDB of alpha-synuclein (as these models have all been built in different cryo-EM maps with different resolutions) - is the same pocket always identified and if not, what is causing that? __We thank the reviewer for this relevant comment. Rather than selecting rotamers or charge states post hoc, we will assess repeatability using corresponding pockets in symmetry-related copies, independent structures belonging to the same structural groups, and repeated experimentally occupied sites. Correspondence will be defined independently of PSI, after which pocket properties and PSI variability will be quantified. We will also examine whether pocket number, volume or similarity depend on structural/model quality.

    3. __ The manuscript implies that similar pockets will bind to similar ligands which makes sense. However it would be helpful to describe that with specifics: i.e. electrostatics, solvent accessibility, water molecules nearby, induced fit etc. An overall conclusion Figure that describes pocket architecture would be useful. __We will explicitly describe the geometric and physicochemical descriptors used in the pocket representation and distinguish these from factors not currently captured, such as detailed hydration, induced fit and ligand–ligand interactions. We will also add a conceptual figure summarizing the structural features represented by the pocketome and those that remain outside its scope. Minor comments:

    4. __ There is no reference to Figure 1 in the main text. Please include MK6240 as well as a tau pet tracer.__ We will add the missing reference to Figure 1 and incorporate MK-6240 and relevant tau PET tracers into the introductory discussion. We will also correct the terminology concerning the PET tracers, including the spelling of florbetaben and the characterization of ACI-12589.

    5. __ Have the authors tried to group filaments based on Scheres amyloid packing algorithm (APD)? __We agree that amyloid packing difference provides a useful orthogonal structural comparison. We will therefore examine the relationship between pocket similarity and APD where technically appropriate, while retaining the existing structural grouping for redundancy control and representative selection rather than replacing the entire classification procedure.

    6. __ Figure 5 legend: L298 Please add that this pathological phenotype (i.e. the inclusions formed in cell and mouse models) do not necessarily recapitulate what is observed in disease. This may or may not be due to the structures formed in these model systems. __We will clarify that pathological phenotypes and inclusions observed in cellular or mouse models do not necessarily reproduce the structures observed in human disease and that our structural analysis should not be interpreted as establishing such equivalence.

    7. __ L390 - Implications for in vitro models: It should be mentioned that it depends on the question. Distinct structures of alpha-synuclein form in synucleinopathies, as such, studying the disease (e.g. in mouse), it is important to study this within that context. Indeed if a binding site is identical, which they are in many of the greek-key like fold of alpha synculein then yes this can be used in the development of a ligand. However, it should always be validated with brain derived filaments.__ We will revise this section to emphasize that the relevance of an in vitro fibril depends on the biological question. In particular, a conserved local binding environment may support the use of an in vitro fibril as a ligand-development model, but this does not establish that the complete structure reproduces the disease-associated filament. We will therefore emphasize the importance of validation against brain-derived filaments when these are available.

    8. __ L408: Missing references to the filament structures. __The relevant filament structures and corresponding references will be added at this location and throughout the manuscript where appropriate. Reviewer #2

    __1, A clear definition of a pocket should be included. How deep can it be, what volume, how solvent accessible? This was not at all clear to me and as shown in Figure S15c, if a larger definition of a pocket is used amyloid-specific ligands are found. This might be expected: small pockets are more likely to be shared in common compared with larger ones. Hence, how does the analysis perform if pockets of larger size are considered? Such an analysis would really improve the article and be of immense use for the field. __We agree that the operational definition of a pocket should be made substantially clearer. The revised Methods will provide the VolSite parameters, volume and accessibility criteria, layer and interface filters, symmetry rule, descriptor definitions and complete filter attrition. We will also perform a sensitivity analysis of the pocket-size criterion using the existing detections and examine how pocket counts and similarity patterns change with pocket size. The 38 ligand-bound structures will provide an independent empirical benchmark for determining whether experimentally occupied sites are represented by the retained pockets or by sites excluded by the current filters.

    __ Abeta was considered in the work, yet there is little discussion of its pockets in the latter parts of the Results section. This fibril type is especially interesting as it does not have a fuzzy coat. Please add this detail. __We will expand the comparison of amyloid-β pockets and explicitly discuss how its structural organization differs from the other two proteins considered in the reference dataset. __ Regarding the fuzzy coat, this could occlude some of the pockets. Have the authors considered this fact? A pocket may not be solvent exposed in the context of the full fibril and not just the fibril core. __We agree that the fuzzy coat may influence the biological accessibility of pockets. Our analysis is based on the experimentally resolved ordered fibril core and therefore does not reconstruct unresolved or dynamically disordered regions. We will make this limitation explicit and distinguish accessibility in the resolved structural model from accessibility in the complete biological fibril. __ A unique feature of the amyloid fold is its repeating beta strands and the twist. So how does this impact a pocket? Surely many pockets will repeat along the fibril axis creating grooves rather than pockets? Please explain. And regarding the twist, how does this affect the pockets defined? The polymorph analysis in Calypso uses only a very few layers so the twist is not considered in their analysis. __We will clarify that our pocket representation describes local environments within a finite number of fibril layers and does not necessarily capture the complete recognition surface of a ligand spanning several rungs. This limitation will also be discussed in relation to the ligand-bound validation analysis. __ As the field will be especially interested in finding polymorph-specific ligands for disease related amyloids, I would appreciate adding a section that compare the pockets in those fibrils for Abeta, Alpha-synuclein and tau with detailed figures to assist the analysis and clarity. __We will strengthen the cross-protein comparison of amyloid-β, tau and α-synuclein, including representative examples of recurrent and more restricted pocket environments. We will also make the scope explicit throughout the manuscript so that conclusions concerning shared environments are clearly understood as applying to the three-protein reference dataset unless independently supported by the external validation analysis.

    Reviewer #3

    1. __ The filters may exclude experimentally observed ligand-binding site classes . __We agree that the current filters may exclude experimentally occupied site classes and that this limits the interpretation of the original conclusions. We have therefore initiated an analysis of 38 ligand-bound fibril structures in which ligands are removed before pocket detection, allowing us to determine whether experimentally occupied sites are detected and, if not, which filter excludes them. We will also distinguish surface, enclosed and interface-related site classes in the revised analysis. Importantly, we will revise the conclusions so that the reported recurrence of pockets applies to the sampled ordered-core surface pocket class and is not presented as evidence that selectiv sites are generally rare across all amyloid ligand-binding modes.
    2. __ The dataset is restricted to three proteins without justification . __We agree that the restriction to amyloid-β, tau and α-synuclein needed to be stated more explicitly. The revised manuscript will clearly define these three proteins as the reference dataset and will qualify the corresponding conclusions accordingly. To test transferability without compromising the independence of the validation, we have additionally completed representative selection and pocket detection for TDP-43, TMEM106B and transthyretin. These proteins will be treated as an external hold-out set and projected into the frozen three-protein reference space rather than being used to redefine the reference metric.
    3. __ The central claim is not yet supported by an explicit test. __We will quantify within- and between-protein similarity using structure-balanced analyses and blocked/hierarchical resampling, accounting for the fact that multiple pockets can originate from the same structure. We will compare the observed mixing with appropriate structure-level null models and report both pocket-weighted and structure-balanced results.
    4. __ The metric is unvalidated and the counts behind it are unreported__ __Major comment 4a : __We agree that the PSI scale requires empirical calibration. We will report the full and nearest-neighbour distance/PSI distributions and interpret individual values relative to these empirical distributions rather than treating a fixed PSI threshold as equivalent to uniqueness. We will also explicitly state which dataset defines the reference scale and retain this scale unchanged when projecting the external validation datasets.

    Major comment 4b : We will provide the exact mathematical definition of the descriptor vector and distance calculation, including the descriptor index set, units, preprocessing and scaling. We will additionally assess the contribution of descriptor families and test whether the principal conclusions are robust to standardization and feature-family sensitivity analyses.

    Major comment 4c : We will quantify similarity and variability for symmetry-related pocket copies, closely related independent structures and repeated experimentally occupied sites, defining correspondence independently of PSI. We will report dispersion and matching failures rather than relying only on average similarity.

    Major comment 4d : We will add a complete structure-to-pocket flow showing the number of structures at each stage, the structural groups, representatives, raw cavities, filter attrition and final retained pockets. We will explicitly distinguish the approximately 400 structures used for structural classification from the representative structures subjected to the pocket analysis and from the final pocket dataset.

    __Major comment 4e : __We agree that the use of one representative per structural group limits the statistical power to establish how frequent polymorph-specific pockets are. We will therefore qualify the absence of polymorph-specific clusters as an observation within the current representative sampling rather than evidence that such environments are intrinsically rare.

    __Major comment 4f : __We agree that model quality is particularly relevant because the pocket descriptors depend on side-chain placement. We will report resolution and available model-quality information for the representatives and test whether pocket properties and similarity correlate with structural quality. We will also perform targeted quality sensitivity analyses and examine the effect of retaining non-identical second protofilaments where appropriate.

  2. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary

    The authors ask whether amyloid fibril polymorphism produces enough diversity in local surface pockets to allow protein-selective or polymorph-selective ligands. They treat the fibril surface as a catalogue of cavities rather than a set of folds. From 400 cryo-EM structures of α-synuclein, tau and amyloid-β they keep one representative per fold group, detect surface cavities, filter to those judged accessible to a small molecule, encode each as a vector of geometric and physicochemical descriptors, and compare all pairs with a single similarity index. The resulting map is the amyloid pocketome.

    They find it largely continuous. Pockets recur across polymorphs and across the three proteins, they do not track fold classification, and the commonest class is a small cavity formed by a few charged or polar side chains. The Concluding perspective states the objective plainly: the work "argues for a shift from a fold-centric to pocket-centric paradigm in amyloid ligand discovery" (l. 480), reports that "many pockets are shared across proteins and polymorphs, whereas truly isolated pockets are rare" (l. 483-485), and holds that this landscape "helps explain why selective amyloid ligands have been difficult to obtain" and "defines the structural limits and opportunities for selective amyloid targeting" (l. 487-492). The study is computational throughout: no new structures and no binding measurements are reported.

    Major comments

    1. The filters may exclude experimentally observed ligand-binding site classes Ligand-bound amyloid structures include at least two recurrent geometries, defined by the orientation of the ligand relative to the helical axis. The present filters appear to retain exposed surface-groove sites while removing several enclosed or interface-bound sites. Perpendicular (columnar). The ligand lies across the fibril and copies stack into a column along the axis. MK-6240 adopts "a stacked arrangement perpendicular to the fibril axis" and spans about two tau rungs [R2]; GTP-1 stacks across three rungs [R1]; flortaucipir sits at about 46{degree sign} to the helical axis [R3] and F0502B at about 50{degree sign} [R6]. Burial is substantially ligand-ligand: 243 Ų against 208 Ų of protein contact for MK-6240 [R2]. EGCG is the flat limit of this mode, in register with the 4.7 to 4.8 Šrise at 1:1 stoichiometry [R5]; separating flat from tilted cases is secondary to the point made here. Parallel. The ligand's long axis runs along the fibril axis, lying lengthwise in a surface groove and spanning several rungs. APN-1607 binds tau this way at sites 1, 2a and 2b, "parallel to the long helical axis", in both paired helical and straight filaments [R4]. PM-PBB3 is modelled the same way on TMEM106B, "paralleling to fibril axis and spanning four rungs" [R12]. Filter 3 removes cavities at the protofilament interface, and filter 4, together with the criterion illustrated in Fig. S9E for pockets "localized within the adopted monomeric fold", removes cavities buried within the fibril or enclosed by a single fold. Several perpendicular or cleft-bound ligands occupy these site classes. MK-6240 and APN-1607 site 3 are resolved within the C-shaped cavity of the tau fold [R2, R4]; EGCG binds tau at "the polar cleft at the intersection of the two protofilaments" [R5]; F0502B occupies a site at "the protofilamental interface of WT polymorph 5a" on 8ZMY [R7], one of the manuscript's own representatives. What survives filtering appears dominated by shallow surface grooves and by the recurrent 100 to 150 ų cavities defined by two or three residues that the paper calls non-discriminatory (l. 216-222; l. 269-271). State whether the pipeline detects and retains the C-shaped cavity of the Alzheimer tau fold and, if not, which filter removes it. Under either objective stated in the Concluding perspective, this materially limits the claimed scope. If the aim is to describe the amyloid pocketome, a map that omits empirically occupied site classes is incomplete. If the aim is to guide selective ligand design, excluding sites used by several of the best structurally characterised tracers leaves the central question only partially addressed. Retain the excluded classes as separate strata and analyse them alongside the retained surface grooves rather than discarding them before analysis. Isolated pockets are then counted in what remains. Concluding from that count that selective sites are rare in general (l. 385), and that this explains why selective ligands have been hard to obtain (l. 487-488; also l. 355-360, l. 414-419), selects on the predictor. Fold-restricted sites do exist: flortaucipir gives clear density on chronic traumatic encephalopathy Type I filaments while the same study was "unable to visualize additional cryo-EM density for flortaucipir for AD paired helical or straight filaments" [R3]. Fig. S10 does not settle this. VolSite finds a 511 ų cavity coincident with F0502B in 7WMM, so detection works. But 7WMM is holo, so a cavity recovered after deleting the ligand is a cast of it; the test is whether it appears in the apo form, since the representatives are largely apo. And 511 ų is set partly by the four-layer model and the tip filter, so comparing it with a 100 to 150 ų per-rung trough in one unscaled Euclidean space compares different objects. Requested (essential; existing data). Perform a sensitivity analysis of the pocket definition rather than treating the 80 ų minimum (l. 524) as sufficient evidence of ligand relevance. The one cavity shown here to accommodate a real fibril ligand is 511 ų (Fig. S10), whereas the recurrent class called non-discriminatory is 100 to 150 ų (l. 217-218) and cross-amyloid cluster 4 is "approximately 150 ų or less" (l. 266-267). Because a ligand may engage a shallow subpocket or extend beyond the detected volume, this example does not establish one correct cutoff. Re-run the analysis across a justified range anchored to the dimensions and contact footprints of published ligand sites; report how the pocketome, intermingling and isolated-pocket counts change; and report buried contact area per pocket alongside volume. Also compare apo and holo forms of the same polymorph. Separately, recast the conclusion as an explanation of ligand promiscuity rather than of the absence of selectivity, scope the negative-design proposal to the site classes actually sampled, and qualify or remove l. 355-360, l. 385 and l. 487-488. Report how many cavities each filter removes, per protein, and reinstate interface and enclosed classes as separate strata so that empirically occupied binding-site classes are represented. The expected outcome is itself the interesting result. A site able to host a ligand across several rungs is an extended groove or cleft, and not every fold presents one: the C-shaped cleft of the Alzheimer paired helical filament does, whereas flatter folds such as the Pick's disease filament (6GX5, a representative here) are unlikely to. If raising the threshold leaves pockets concentrated on a subset of folds, that is fold-level discrimination emerging from the authors' own data, and it bears directly on the conclusion drawn at l. 385.
    2. The dataset is restricted to three proteins without justification The Methods inclusion criteria (l. 497-501) specify cryo-EM structures from the Amyloid Atlas and exclude ssNMR assemblies, monomers and peptides. They never state a restriction to amyloid-β, tau and α-synuclein. The three proteins first appear at the classification step (l. 504-506) as an assumption, so the criteria as written do not reproduce the dataset, and the restriction is absent from the limitations at l. 441-470. The Atlas already holds cryo-EM fibril structures for TDP-43, TMEM106B, transthyretin, IAPP, serum amyloid A, β2-microglobulin, immunoglobulin light chain and prion protein. Extending the map would be a direct test of the central claim, although it would require additional curation, representative selection and filtering rather than being purely mechanical. If convergence reflects backbone-lined grooves that any cross-β spine presents, these folds should intermingle with the three already included; if the map instead separates once more proteins are present, the claim changes. The restriction also sets the scale of the whole metric, since σ is the standard deviation of distances within this three-protein set (l. 543-544). "Isolated in the amyloid pocketome" therefore currently means "isolated among these three proteins", while the title, the term "cross-amyloid pockets" (l. 265) and the Concluding perspective generalise to amyloid as such. TMEM106B is the most pointed omission. It forms amyloid filaments in aged and diseased human brain [R9], and it is a documented off-target of tau PET tracers: both [¹⁸F]PM-PBB3 and [¹⁸F]flortaucipir bind TMEM106B-containing choroid plexus homogenate with high affinity, and PM-PBB3 co-localises with TMEM106B-immunoreactive Biondi ring structures in the choroid plexus epithelium [R10]. Cross-protein binding by clinical tracers is the phenomenon this manuscript sets out to explain, in a protein it excludes. The comparative literature the manuscript relies on is in any case already broader than three proteins: ref. 78 compares disease-associated folds of prion protein, tau, α-synuclein, TDP-43 and TAF15 [R8]. Requested (essential scope correction; expansion strongly recommended). State and justify the three-protein restriction in Methods and add it to the limitations. Revise the title, Abstract and conclusions so that the present map is explicitly a pocketome of the sampled site classes in amyloid-β, tau and α-synuclein, or broaden the analysis sufficiently to support the general terminology. A strongly recommended extension is to add other Atlas proteins, at minimum TDP-43, TMEM106B and transthyretin, and report whether intermingling, the σ-normalised outlier set and the clusters change. If distinguishing sites emerge, describe them and outline how they might be targeted.
    3. The central claim is not yet supported by an explicit test "Extensively intermingled" (l. 160-162) and "did not systematically co-localize" (l. 183-184) are inferred from the colours in Figs. 5, 6 and S11 to S13; no test is reported. A minimum spanning tree connects every node by construction (l. 549-551), so connectivity itself cannot establish mixing or non-isolation. A terminal node joined by a long edge can still represent geometric isolation, but the figures provide no interpretable PSI or edge-length scale. Fig. 5 also shows long single-colour runs, which the authors concede as "local enrichments" (l. 162-163). Requested. Compare PSI within proteins against PSI between proteins using structure-balanced summaries. Count how often neighbouring pockets share a label and compare the result with a null obtained by permuting labels at the representative-structure or fold level, preserving all pockets from the same structure. Several pockets are nested within one model, and every pocket contributes to many pairwise PSI values; pocket-level shuffling or treating all PSI pairs as independent would inflate the effective sample size and create impossible mixed labels within a structure. Use blocked permutations or a hierarchical bootstrap with uncertainty intervals, report both pocket-weighted and structure-balanced estimates, and show that the result is not driven by structures yielding unusually many cavities. Repeat for polymorph and for fibril source, which Fig. 2B records but no analysis uses. The second claim, that pockets are decoupled from fold, compares two rulers that measure different things. Pockets are defined by side chains (l. 565-566); polymorph groups come from a Cα RMSD (l. 506-508), which is blind to side-chain packing, since Cα RMSDs "can lead to relatively low values for structures that share similar backbone conformations but differ in their side-chain packing interactions" [R8]. Two structures can therefore share a group and still present different pockets, which is precisely the reported observation (l. 185-186). Requested. Drop the hand-cut groups and correlate PSI against pairwise fold distance directly, computed both as Cα RMSD and as the amyloid packing difference [R8], which counts differing side-chain packing contacts instead of superposing coordinates. If PSI tracks the packing difference but not Cα RMSD, the decoupling is an artefact of the comparator. RMSD also assumes a common superposition, and the α-synuclein models do not share one: ordered spans run 42 to 102 residues and about a quarter of entries contain internal chain breaks (Fig. S1), so equal RMSD values do not describe equal differences. Tau is largely exempt (Fig. S2); amyloid-β mixes Aβ40 and Aβ42 constructs on one axis (Fig. S3). Requested. State the superposed residue range for every comparison, and report whether the α-synuclein and amyloid-β groupings survive restriction to a common core.
    4. The metric is unvalidated and the counts behind it are unreported (a) No scale. PSI = exp(−d²/2σ²) restates distance in units of σ: 0.624 is 0.97σ, 0.148 is 1.95σ, and 0.005 is 3.3σ. No distribution is shown, so no value can be judged typical or extreme. The kernel strongly compresses the tail beyond about 3σ, and the "PSI below 0.005" rule operationally groups a broad range of large distances into the same outlier category. It therefore does not by itself separate unusual from unique, yet it defines the selective-targeting opportunities (l. 277-281; l. 320-325). σ is computed within the dataset analysed (l. 543-544), so global and protein-specific values sit on different scales while being quoted together, and every added structure changes every PSI, complicating the claim that new structures can be placed in an unchanged existing map (l. 464-470). Requested. Publish the full and nearest-neighbour PSI distributions and define "similar" and "isolated" against them; rank outliers on raw distance in σ; state which matrix each quoted value comes from; and either give a dataset-independent normalisation or drop the extensibility claim. (b) Ambiguous distance, unscaled descriptors. "Non-zero descriptor values only" (l. 540-541) names no index set. Dropping globally zero columns, or keeping descriptors non-zero in either pocket, equals the full-vector distance; keeping only those non-zero in both compares each pair in a subspace of different size and discards the largest differences, since a descriptor present in one pocket and absent in the other is deleted rather than counted (Fig. 4D: OD1 = 32.2 against 0). Small simple cavities carry the most zeros, so this rule could bias the analysis toward the reported convergence. No scaling is specified, so volume in ų shares a sum with bounded percentages and may dominate. Requested. Give the rule as one equation, recompute under the standardized full-vector alternative and report the correlation between PSI and volume difference. With 109 descriptors and only on the order of 10² retained pockets, correlated descriptor families can overweight one physical property and Euclidean distances can concentrate. Report descriptor correlations or effective rank and the nearest-to-farthest distance contrast, and show that nearest neighbours and DBSCAN assignments are stable after standardization, correlation pruning or PCA, and feature-family ablation. (c) No repeatability benchmark. Two controls sit in the authors' own data. Symmetry-related copies of one pocket are detected twice and the smaller discarded (l. 532-534); their similarity distribution would provide an empirical repeatability envelope or expected upper range, and Fig. S9F already shows five clouds against four on supposedly identical protofilaments. Second, structures sharing a fold group at different resolutions can be run against each other to test whether the same surface yields the same pockets. Until one is reported, observed PSI differences cannot be separated from detection and model variability. Requested. Report both controls, including matching failures and dispersion rather than only a mean PSI. (d) Counts. The Abstract and Significance statement attribute the analysis to 400 structures (l. 5-6; l. 21-22), but pocket detection is performed on roughly 50 representatives selected after classification of those structures. No pocket total is given, in any breakdown, so the DBSCAN result and the claim that "most pockets were not assigned to any cluster" (l. 249-251) cannot be checked. The two headline families are seven pockets from six tau structures and only three pockets spanning two proteins, so the cross-protein example is very small. Requested. Add a flow table from retrieved structures through grouping, representative selection and each filter to retained pockets, and distinguish clearly between the approximately 400 structures used for classification and the approximately 50 representatives subjected to pocket analysis. (e) Limited power for polymorph-specific clusters. No cluster composed only of pockets from a single polymorph was found (l. 275-276), but only one representative structure per group entered the analysis. Several pockets from that representative could in principle form a cluster, but the design supplies no independent within-polymorph replication and has little power to establish that polymorph-specific environments are rare. Requested. Qualify the conclusion as a limitation of the sampling, or test several members of each group. (f) Model quality. Selecting the best-resolved member of each group (l. 125) removes redundancy and is sensible, but it is a relative criterion with no quality floor: a group whose best member is poor still contributes a poor model. The α-synuclein representatives span 1.93 Å (9EUU) to 4.8 Å (9D5C), with 7L7H at 4.0 Å. Global resolution is not identical to local side-chain certainty, but these values raise a material concern for a side-chain-based comparison. One such model carries a reported result: the Lewy-fold match at PSI 0.307 (Fig. 7B) compares a pocket from 9D5C at 4.8 Å against one from 8A9L at 2.2 Å. Separately, five of the 18 α-synuclein representatives contain a second non-identical protofilament, including 6XYO, 8ZMY and 9OBP, so keeping unit 1 only discards a different surface rather than a redundant copy. Requested. Report the global resolution and, where available, local map and model-validation measures for every representative; test whether pocket count, volume and PSI correlate with model quality; apply a justified quality sensitivity analysis; and include non-identical second protofilaments where they were dropped. Bearing on the Concluding perspective (l. 478-492) Read against the analysis, one of the claims made there survives intact.
    • "Ligand selectivity is ultimately governed by the local binding environments exposed on fibril surfaces" (l. 482-483) is the premise rather than a result, and it sits awkwardly with the paper's own material: for a stacked ligand most of the buried surface is ligand-ligand rather than protein, 243 Ų against 208 Ų for MK-6240 [R2], so selectivity there is not governed by the local protein environment alone.
    • "Many pockets are shared across proteins and polymorphs, whereas truly isolated pockets are rare" (l. 483-485) rests on an untested visual reading (point 3), a rarity criterion that saturates and so cannot separate unusual from unique (point 4a), a σ fixed by the three-protein dataset (point 2), and a filtered set that excludes the enclosed sites where ligands are actually observed (point 1).
    • "This constrained pocket landscape helps explain why selective amyloid ligands have been difficult to obtain" (l. 487-488) is not supported. No binding or selectivity measurement enters the analysis at any point, and the ligands being explained bind by a mode the Limitations concede the descriptors do not capture (l. 449-454).
    • "Avoid recurrent cross-amyloid pockets that are structurally predisposed to off-target recognition" (l. 490-491) rests on a cross-amyloid cluster of only three pockets spanning two proteins. Those cavities are 100 to 150 ų, well below the 511 ų of the one site shown here to accommodate a ligand, and whether they constitute complete ligand-binding sites has not been established.
    • "Use in vitro fibrils when they reproduce the relevant local binding environment" (l. 489-490) is the recommendation the analysis does support, and the refinements listed under Patient-derived pockets would strengthen it further. Accordingly, "defines the structural limits and opportunities for selective amyloid targeting" (l. 492) overstates what was done. "Maps the cavity landscape of the sampled site class" is defensible on the present analysis, and would still be a useful contribution. A constructive extension (OPTIONAL): two catalogues this analysis could deliver As run, the analysis chiefly produces a negative result: shared pockets are common and isolated pockets are rare. Two concise catalogues would make the framework more useful and give each result a testable prediction. Broadening them beyond the present three proteins would strengthen their scope but is an optional extension rather than a prerequisite for a properly scoped paper. A catalogue of genuinely selective sites. Rank pockets by similarity to the nearest pocket from another fold. Candidates whose nearest out-of-fold neighbour is distant should be reported with lining residues, volume, accessibility, putative binding mode and the folds in which no counterpart was found. The prediction is direct: ligands designed against these sites should discriminate among a defined fibril panel. A catalogue of minimally selective sites. The converse list is also useful and is closer to what the present data support. Pockets recurring across many folds or proteins are candidate targets for pan-amyloid, pan-tauopathy or other coverage-oriented applications. Report each recurrent class and the proteins and folds in which it occurs; here recurrence is a design specification rather than only a liability. Anchoring makes both lists interpretable. Run published amyloid-ligand cryo-EM complexes through the same pipeline after ligand removal and use the recovered sites as empirical anchors. These include enclosed tau-fold cavities (MK-6240 [R2], flortaucipir [R3]), extended grooves (APN-1607 [R4], F0502B [R6]) and a polar protofilament cleft (EGCG [R5]). If sites occupied by related chemotypes are neighbours, the map gains predictive content; if not, the descriptor set requires revision. A pass/fail test is available now. GTP-1 and MK-6240 independently occupy the same Alzheimer tau site involving Gln351, Lys353, Asp358 and Ile360 [R1, R2]. This provides two checks with known structural answers: whether the pipeline detects and retains that site, and whether it separates the site from tau folds that do not present the cavity, including Pick's disease (6GX5), corticobasal degeneration (6TJX) and progressive supranuclear palsy (7P65). Recovery and discrimination on this case would be a much stronger validation than the single holo example in Fig. S10. One list says where to aim for specificity, the other where to aim for coverage. Both follow from work already done, and together they would convert the pocketome from a catalogue of cavities into an instrument for choosing targets, which is closer to what the Concluding perspective claims.

    Minor comments

    Patient-derived pockets

    • The comparison of patient-derived and in vitro pockets (l. 304-335; Fig. 7) already reports both matches and non-matches, which the surrounding text undersells. Two MSA pockets are matched to in vitro fibrils at PSI 0.624 and 0.335, three Lewy-fold pockets to other polymorphs at 0.154, 0.148 and 0.307, and two Lewy-fold cavities are reported as having no close neighbour at 0.005 or below, with the buried and unassigned-density caveats stated by the authors. The metric therefore produces non-matches as well as matches.
    • Three refinements would let the section carry the weight the Discussion places on it. Report all pockets of every ex vivo structure with ranked nearest in vitro matches rather than a selection. Interpret the values against the background distribution requested in point 4a, without which 0.624 and 0.148 cannot be ranked against one another. And use F0502B as a positive control: its site is in the dataset twice (7WMM, Fig. S10; 8ZMY, a representative) and it is roughly tenfold selective for α-synuclein over tau and amyloid-β by direct affinity measurement [R6], so recovering it as selective by PSI would validate the framework on a ligand of known behaviour.
    • State how MSA's two protofilaments were handled. The text discusses only "protofilament IA" (l. 310) while Figs. S6 and S8 state the representative unit "were always number 1". Published packing differences for this comparison are 8 to 11% for one protofilament against 60% for the other [R8], so the answer depends on which is used.
    • Pocket detection is validated on one example, with no overlap metric and no denominator (l. 141-143; Fig. S10). Tabulate the published amyloid-ligand cryo-EM complexes: cavity detected at the ligand site, numerical overlap, and survival of each filter. The set covers tau [R1-R5], α-synuclein [R6, R7] and TMEM106B [R12]. Internal consistency
    • Fig. 6 caption says the legend shows pocket volume; the in-figure legend reads "Polymorph" with entries "-" and 1.0 to 11.0.
    • Text gives "7V4C pocket 7" at PSI 0.624 (l. 313); the Fig. 7 legend gives pocket 9. The second MSA pocket at l. 314-316 does not appear in Fig. 7A.
    • "Approximately 15 groups per amyloid family" (l. 123) against 18 (Fig. 3), 14 (Fig. S4) and 20 (Fig. S5).
    • The dashed cut in Fig. 3 measures 6.72 {plus minus} 0.02 Å against the stated 5.3 Å (l. 507), with 19 to 20 branches crossing rather than 18. The lines in Figs. S4 and S5 measure 4.80 and 3.31 Å and match their legends. Regenerate Fig. 3 or correct the threshold, and confirm which value the 18-group set used.
    • Fig. S7 shows 13 tau representatives against 14 groups in Fig. S4; the missing one is explained only in the S4 legend, where 9GG6 is merged with 9GG0 and 6GX5.
    • Polymorph labels reach 11 for α-synuclein and 9 for tau, with gaps at amyloid-β 3 and 6, against 18/14/20 groups, plus an unexplained "-" category. Explain the mapping and what "-" denotes.
    • Fig. 4C says six filters; Methods number five plus a symmetry rule (l. 523-534). Fig. S15C is titled "Polymorph-specific pockets" although none were found. Typographic: duplicate titles on the title page, "polymorphs.." (l. 329), "Tha alpha-synuclein pocketome" (Fig. S14), "fibil" (Fig. S10).

    Figures

    • The main and supplementary split does not match where the information is. Fig. 1 contains no data. Fig. 3 shows only the α-synuclein dendrogram, so 55% of structures and 65% of groups are classified in the supplement, and the one dendrogram in the main text is the one that disagrees with its legend. Three supplementary figures are load-bearing: S9, the only place the six filters are shown; S14, on which the isolated-pockets claim rests; S15, the only view of the pocket families. Move Fig. 1 to the supplement, combine the three dendrograms into one main figure, and promote condensed forms of S9, S14 and S15. Fig. S15's image and legend are on separate pages.
    • Figs. 5, 6 and S11 to S14 are screenshots of the interactive TMAP page, with an HTML dropdown visible in the legend box, and Fig. 5 labels proteins by PFAM name ("Tubulin-binding"). Regenerate as vector figures with a PSI scale. The 29-category palette of Fig. S11 is unreadable. A volume-coloured pocketome should be shown, given point 4b.
    • The most useful missing figure: a PSI distribution with internal anchors on one axis (symmetry-mate duplicates, within-structure pairs, cross-polymorph pairs, cross-protein pairs, permuted null), beside the mixing statistic against that null. If two copies of the same pocket do not score near 1, the labels in Fig. 7 cannot be interpreted.
    • Fig. 1: "[¹⁸F]Florbetapen" should read florbetaben. "Clinically-validated" overstates [¹⁸F]ACI-12589, which separates MSA from controls, PD and DLB but shows no increased retention in sporadic PD or DLB [R11]. Second-generation tau tracers are absent although MK-6240 and APN-1607 are central to point 1.

    Methods and reproducibility

    Not reproducible as written. Absent: VolSite parameters (given as values "previously validated for protein-protein interaction cavities", l. 521-522); DBSCAN eps and min_samples, the definition of the PSI-derived distance matrix, and cluster and noise counts, noting that Fig. S14A shows one chained cluster holding roughly 54 of about 75 assigned pockets across every arm of the tree; the 109 descriptors with units and scaling; MOE version and QuickPrep settings, including pH and whether the default minimisation was disabled, since l. 515 says none was applied; how the four layers were chosen; whether ligands, ions and waters were removed before detection and how many holo structures contribute, since 8ZMY and 9QYL are ligand- or cofactor-bound; TMAP parameters and software versions; the 400 PDB IDs with retrieval date; and analysis code. Supplementary Files 1 to 3 are announced on SI p. 1 but are absent from the supplied review packet; provide them and describe their contents. Only two of the six filters are stated operationally, volume and layer position; give a numeric criterion and per-filter attrition count for each. Manual interventions appear only in SI legends. "Repeated along the fibril axis" (l. 134-135; l. 384; l. 445) is never measured. Useful sensitivity analyses include the RMSD cut, layer count, representative choice, and inclusion versus exclusion of interface and buried pockets. Prior literature Ref. 78 is cited as a preprint but is published [R8]; check refs. 77 and 80 likewise. The ligand-bound amyloid literature is otherwise absent although it bears on point 1: the published classification of binding geometries on α-synuclein polymorphs [R7] is uncited although one of its structures is used as a representative, and no ligand-bound tau or amyloid-β structure is discussed. The FibrilSite preprint (ref. 80) reaches the opposite conclusion on cross-protein site sharing and is answered at l. 421-427 only by appeal to observed ligand non-selectivity, which this manuscript does not test; compare the two methods on a shared subset of α-synuclein structures.

    References

    [R1] Merz GE, et al. Nat Commun 14:3048 (2023). GTP-1; PDB 8FUG.

    [R2] Kunach P, et al. Nat Commun 15:8497 (2024). MK-6240; PDB 8UQ7.

    [R3] Shi Y, et al. J Mol Biol 435:168025 (2023). Flortaucipir, CTE filaments; PDB 8BYN.

    [R4] Shi Y, et al. Acta Neuropathol 141:697-708 (2021). APN-1607; PDB 7NRV, 7NRX.

    [R5] Seidler PM, et al. Nat Commun 13:5451 (2022). EGCG on tau; PDB 7UPG.

    [R6] Xiang J, et al. Cell 186:3350-3367 (2023). F0502B; PDB 7WMM.

    [R7] Liu K, et al. Proc Natl Acad Sci USA 121:e2321633121 (2024). Binding geometries on α-synuclein polymorphs; PDB 8ZMY, 8X7B, 8X7O, 8X7L, 8X7Q, 8ZLI.

    [R8] Scheres SHW. Structure 34:1061-1071 (2026). Amyloid packing difference.

    [R9] Schweighauser M, et al. Nature 605:310-314 (2022). TMEM106B filaments.

    [R10] Yokoyama Y, Harada R, Kudo K, et al. Transmembrane protein 106B amyloid is a potential off-target molecule of tau PET tracers in the choroid plexus. Nucl Med Biol 142-143:108986 (2025). Postmortem autoradiography and binding assays; PM-PBB3 co-localises with TMEM106B-immunoreactive Biondi ring structures, and both PM-PBB3 and flortaucipir bind TMEM106B-containing choroid plexus homogenate with high affinity.

    [R11] Smith R, et al. Nat Commun 14:6750 (2023). [¹⁸F]ACI-12589.

    [R12] Zhao Q, et al. Cell Discov 10:50 (2024). PM-PBB3 on TMEM106B; PDB 8J7N. Density weak and also present in apo and PiB maps; pose low-confidence. APN-1607, PM-PBB3 and florzolotau are one compound; flortaucipir is AV-1451 or T807.

    Significance

    General assessment

    The framing is the strongest aspect. The gap between the growth of amyloid cryo-EM structures and the absence of selective ligands is real, and locating the bottleneck at pocket discriminability rather than fold classification is productive. The distinction between a pocket being ligandable and being discriminable is useful, and using a pocketome as a negative filter, to exclude sites where selectivity is implausible before chemistry is invested, is the most valuable idea here. The filtering logic reflects real thought about what a fibril-bound ligand can reach, the limitations section is candid, and the data are deposited. The principal weakness is conceptual. The pocketome pools site classes that the ligand-bound literature treats separately, and because the filters remove interface and enclosed cavities, the reported continuum may be a property of the sampled subpopulation rather than of amyloid surfaces. The work supports the claim that certain ligand classes are promiscuous. It does not establish how rare selective sites are across amyloids, and published fold-discriminating and protein-selective examples constrain any broader negative inference. Beyond that, no hierarchy-aware statistic or repeatability control is reported; the similarity index is under-specified and uncalibrated; and the scope wording blurs the distinction between approximately 400 structures used for classification and approximately 50 representatives used for pocket analysis. Most requested corrections are re-analyses or clearer reporting of data already in hand. One structural feature of the manuscript deserves comment. The Limitations section (l. 441-470) already concedes several of the points on which the central conclusions depend: that detecting a cavity does not demonstrate binding, affinity or selectivity (l. 443-447); that ligands binding "in repeated arrays along the fibril axis" and stabilised by "both fibril-ligand and ligand-ligand interactions" are "not fully captured by pocket descriptors based on local cavity geometry" (l. 449-454); and that the analysis is bounded by the structures available. Each is accurate. The difficulty is that they are offered as caveats on precision while the Discussion draws conclusions that require them to be false. A catalogue of detected cavities is not undermined by the fact that stacking is unrepresented; an explanation of why amyloid ligands are non-selective is, because the ligands in question bind by stacking. Likewise, a map of cavities is unaffected by the caveat that detection is not binding, whereas the claim that isolated pockets are rare, and that selective targeting is therefore structurally constrained, requires detected cavities to stand in for targetable sites. As written, the paper is a legitimate descriptive resource carrying an interpretive layer its own Methods disclaims. Either scope the conclusions to what the analysis supports, or address the limitations rather than acknowledge them.

    Advance

    Conceptual and technical rather than mechanistic or clinical. It reframes amyloid ligand design from fold-centric to pocket-centric and ports pocketome methodology to filamentous assemblies, which required non-trivial decisions about what counts as a pocket on a helical polymer. The closest work is the FibrilSite preprint (ref. 80), which reaches the opposite conclusion on cross-protein site sharing; the framework is complementary to the packing-difference metric [R8]. Against a ligand-bound literature that has characterised sites one structure at a time [R1-R7], a population-level view is new and is the right instrument for the question. The most useful specific contributions are the identification of recurrent Lys/Arg/Gln/Tyr cavities as poorly discriminating, and the demonstration that some patient-derived α-synuclein pockets have in vitro counterparts.

    Audience

    Specialised: amyloid structural biologists working on cryo-EM of tau, α-synuclein and amyloid-β filaments; computational and structure-based design groups working on shallow or interface-like surfaces; and PET tracer programmes in neurodegeneration, for whom the negative-design idea and the in vitro model-selection argument are directly relevant. With the statistics and the geometric stratification in place, the general message, that polymorphism at the fold level need not imply polymorphism at the pocket level, would reach a broader biophysics and drug-discovery readership. The framework transfers to TDP-43, TMEM106B, hnRNPA1/A2 and transthyretin.

    Reviewer expertise

    Cryo-EM of amyloid fibrils and helical reconstruction; structural biology of tau and α-synuclein polymorphs; ligand-bound filament structures and PET tracer binding modes; computational structure-based design; clustering methodology on structural datasets.

  3. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    This manuscript describes an interesting analysis of the presence of absence of binding pockets displayed on the surface of all know amyloid fibril structures of Abeta, tau and alpha-synuclein. Using their definition of a 'pocket' the authors show that fibrils of the same protein type of different polymorph type (defined by the Calypso algorithm developed by Connor et al) or those from different proteins can share common pockets with a few examples of fibril type specific pockets.

    The work will be of broad interest to those seeking to develop amyloid-specific binders for discovery research and for its translation into the clinic in neurodegenerative disorders. I have comments that I hope will improve the understanding and impact of this analysis:

    1. A clear definition of a pocket should be included. How deep can it be, what volume, how solvent accessible? This was not at all clear to me and as shown in Figure S15c, if a larger definition of a pocket is used amyloid-specific ligands are found. This might be expected: small pockets are more likely to be shared in common compared with larger ones. Hence, how does the analysis perform if pockets of larger size are considered? Such an analysis would really improve the article and be of immense use for the field.
    2. Abeta was considered in the work, yet there is little discussion of its pockets in the latter parts of the Results section. This fibril type is especially interesting as it does not have a fuzzy coat. Please add this detail.
    3. Regarding the fuzzy coat, this could occlude some of the pockets. Have the authors considered this fact? A pocket may not be solvent exposed in the context of the full fibril and not just the fibril core.
    4. A unique feature of the amyloid fold is its repeating beta strands and the twist. So how does this impact a pocket? Surely many pockets will repeat along the fibril axis creating grooves rather than pockets? Please explain. And regarding the twist, how does this affect the pockets defined? The polymorph analysis in Calypso uses only a very few layers so the twist is not considered in their analysis.
    5. As the field will be especially interested in finding polymorph-specific ligands for disease related amyloids, I would appreciate adding a section that compare the pockets in those fibrils for Abeta, Alpha-synuclein and tau with detailed figures to assist the analysis and clarity.

    Referee cross-commenting

    All three referees appeared to like the concept of the manuscript and all three ask for more detail and to include more proteins to test the generality of the claims made. I agree with all the suggestions and I hope the authors can address the comments with the requested details and analyses added. If so, this would then make an excellent reference for those working in the amyloid field.

    Significance

    This manuscript describes an interesting analysis of the presence of absence of binding pockets displayed on the surface of all know amyloid fibril structures of Abeta, tau and alpha-synuclein. Using their definition of a 'pocket' the authors show that fibrils of the same protein type of different polymorph type (defined by the Calypso algorithm developed by Connor et al) or those from different proteins can share common pockets with a few examples of fibril type specific pockets.

    The work will be of broad interest to those seeking to develop amyloid-specific binders for discovery research and for its translation into the clinic in neurodegenerative disorders.

  4. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    This manuscript presents a computational framework for characterizing ligand binding pockets for amyloid filaments. They construct a "pocketome" and the authors describe binding site similarities across different amyloid proteins. They identify conserved and polymorph-specific pockets (which is important for specificity of ligands) and discuss implications for selective ligand design.

    The work does address an important problem in that amyloid filaments typically have limited sites in which ligands can bind to and these can be shared across different amyloid polymorphs which is a strength and conceptually interesting. This will be of specialized interest to the ligand-amyloid field. I have a few concerns:

    Major comments:

    1. Although this is a computational study, there is a lack of validation. This paper is mainly focused on alpha-synuclein however it would be useful if the authors could validate on tau where ligands such as APN-1607 and MK-6240 have been found to bind to AD PHFs and SFs by cryo-EM. They could also test AV1451 which binds to CTE filaments by cryo-EM.
    2. Robustness of the pocket detection pipeline. If you change the rotamer of a side chain where the ligand has been observed to bind or the charge state, does the pocket remain stable? I suppose one could test this on the high redundancy of the same in vitro structures that reappear in the PDB of alpha-synuclein (as these models have all been built in different cryo-EM maps with different resolutions) - is the same pocket always identified and if not, what is causing that?
    3. The manuscript implies that similar pockets will bind to similar ligands which makes sense. However it would be helpful to describe that with specifics: i.e. electrostatics, solvent accessibility, water molecules nearby, induced fit etc. An overall conclusion Figure that describes pocket architecture would be useful.

    Minor comments:

    1. There is no reference to Figure 1 in the main text. Please include MK6240 as well as a tau pet tracer.
    2. Have the authors tried to group filaments based on Scheres amyloid packing algorithm (APD)?
    3. Figure 5 legend: L298 Please add that this pathological phenotype (i.e. the inclusions formed in cell and mouse models) do not necessarily recapitulate what is observed in disease. This may or may not be due to the structures formed in these model systems.
    4. L329: typo error with two ".."
    5. L390 - Implications for in vitro models: It should be mentioned that it depends on the question. Distinct structures of alpha-synuclein form in synucleinopathies, as such, studying the disease (e.g. in mouse), it is important to study this within that context. Indeed if a binding site is identical, which they are in many of the greek-key like fold of alpha synculein then yes this can be used in the development of a ligand. However, it should always be validated with brain derived filaments.
    6. L408: Missing references to the filament structures.

    Significance

    This manuscript presents a computational framework for characterizing ligand binding pockets for amyloid filaments. They construct a "pocketome" and the authors describe binding site similarities across different amyloid proteins. They identify conserved and polymorph-specific pockets (which is important for specificity of ligands) and discuss implications for selective ligand design.

    The work does address an important problem in that amyloid filaments typically have limited sites in which ligands can bind to and these can be shared across different amyloid polymorphs which is a strength and conceptually interesting. This will be of specialized interest to the ligand-amyloid field.

  5. Version published to 10.64898/2026.02.25.707901 on bioRxiv