AI-guided discovery of atypical protein assemblies
A scoring framework that treats low-confidence AlphaFold models as signal rather than failure flagged the NRC7 clade as structurally odd, and electron microscopy then showed those proteins assemble into 11-mers instead of hexamers.
Three NRC helpers have been resolved by cryo-EM and all three are hexamers, which invites the assumption that the clade is hexameric. The Structural Novelty Index tests that assumption instead of inheriting it. Eleven parameters, eight from human experts and seven proposed by the co-scientist AI system, score AlphaFold 3 hexamer models on interface quality, ring symmetry and alpha1-helix geometry. Applied to 637 non-redundant NRC proteins from 346 Solanaceae genomes, it put thirteen of eighteen clades in a core cluster and pushed five out. NRC7 had among the lowest confidence scores while still modelling as a ring. Purified autoactive NRC7 from potato gave particles about 250 A across, not the 150 A a hexamer would give, and the 2D averages count eleven protomers. Tomato and N. benthamiana orthologs do the same.
Key findings (10)
The Structural Novelty Index for NRC hexamers (SNI-NRC-Hexa) uses 11 parameters covering model confidence and inter-protomer interfaces (ipTM-LCB, total contacts, buried surface area), ring symmetry (D-apex, rotational sigma, S-proto), MHD-to-P-loop distance, CC-domain hydrophobicity, and alpha1-helix length, angle and amphipathic moment. Eight came from human experts, seven from the co-scientist AI system, with four overlapping.
Evidence framework design combining expert-defined and AI-generated features
Source Figure 1A; figure S1; data S1
AlphaFold 3 models NRC resistosomes accurately. All nine helper predictions in the benchmark aligned to their cryo-EM structures with RMSD below 1 A, modelled as hexamers with 25 oleic acid molecules standing in for the plasma membrane, three replicates per sequence.
Evidence AlphaFold 3 modelling benchmarked against experimental structures (NbNRC2a PDB 9FP6, SlNRC3 PDB 9RI9)
Source Figure 1B; figure S2
The index separates helpers from sensors cleanly. Hierarchical clustering of 18 benchmark models (9 NRC helpers, 9 NRC-S) gave two distinct clusters, exploiting the prior knowledge that helpers form hexamers and their phylogenetically related sensors do not.
Evidence hierarchical clustering of z-score normalised SNI parameters
Source Figure 1A
Applied at scale, the index flags five clades as divergent. 346 Solanaceae genomes from 85 species were reannotated, yielding 197,834 NLR proteins from 15,079,126 total; 2,658 fell in the NRC clade, reduced to 637 non-redundant sequences across 18 phylogenetic clades. Thirteen clades formed a core cluster containing all the empirically resolved structures; five sat outside it, with NRC4u and NRC7 lowest on mean ipTM and markedly reduced inter-protomer contacts.
Evidence genome reannotation, phylogenetics, AlphaFold 3 modelling of 637 sequences in triplicate, penalised per-clade mean scoring, hierarchical clustering
Source Figures 2A and 2B; figures S4; data S3, S4, S5
Low confidence means different things for different clades. NRC4u models did not form resistosome-like assemblies at all, while NRC7 models did form rings despite similarly low ipTM. That distinction is what made NRC7 the candidate worth testing.
Evidence inspection of representative models from the two most divergent clades
Source Figure 2C
Potato NRC7 carrying the autoactivating MHD substitution D498V (StNRC7DV) triggers cell death in N. benthamiana comparable to the autoactive SlNRC3DV control, while wild-type StNRC7 does not.
Evidence transient expression cell death assay, photographed 5 days post infiltration
Source Figures 3A-3C
Purified activated NRC7 forms an 11-protomer ring. StNRC7EEE+DV (MADA mutations L20E/L24E/L28E added to permit purification) gave ring-shaped particles about 250 A in diameter by negative-stain EM, against the roughly 150 A expected for a canonical hexamer. Counting peripheral densities, likely the LRR domains, in the 2D class averages gives 11 protomers.
Evidence immunoprecipitation-electron microscopy workflow; strep-tag affinity purification from N. benthamiana; negative-stain EM and 2D classification
Source Figures 3D and 3E
The 11-mer is a clade property, not a quirk of one protein. SlNRC7 and NbNRC7, representing distinct branches of the clade, both formed 11-mer assemblies as EEE+DV variants under the same workflow.
Evidence negative-stain EM 2D class averages of three orthologs, with hexameric SlNRC3EEE as the outgroup comparison
Source Figure 4
AlphaFold 3 could not model NRC7 at any stoichiometry. Modelling across alternative oligomeric states from 5-mers through 11-mers gave no confident assemblies, including the 11-mer that turns out to be correct. The prediction identified the anomaly without being able to solve it.
Evidence AlphaFold 3 modelling of NRC7 across oligomeric states
Source Figure S7
Neither the human-defined nor the AI-derived parameter set matched the combined index. Each alone produced one benchmark misclassification (NbNRC4c for the human set, Rpa1 for the co-scientist set). On the full 637-protein dataset both flagged NRC7, but the co-scientist parameters gave finer granularity, identifying four unconventional clades against the human set's three proteins outside the hexamer cluster.
Evidence parallel hierarchical clustering with each parameter set independently
Source Figures S5 and S6
NLRs
| Name | Class | Role in this paper |
|---|---|---|
| NRC7 StNRC7, SlNRC7, NbNRC7, StNRC7DV (autoactive), StNRC7EEE+DV | helper subject | central subject; the clade flagged by SNI and shown by negative-stain EM to form 11-mer resistosomes |
| NRC2 NbNRC2a | helper comparator | benchmark hexamer with an experimental structure (PDB 9FP6) |
| NRC3 SlNRC3, SlNRC3DV, SlNRC3EEE | helper comparator | benchmark hexamer (PDB 9RI9); autoactive positive control for cell death; hexameric outgroup in the EM comparison |
| NRC4 NbNRC4c, NtNRC4, LbNRC4 | helper comparator | benchmark hexamer; NbNRC4c is the misclassification produced by the human-defined parameter set alone |
| NRC4u | helper subject | the other lowest-confidence clade; contrasted with NRC7 because it forms no ring, and left untested experimentally |
| NRC6 | helper comparator | the near neighbour that models as a canonical hexamer, sharpening the NRC7 anomaly |
| NRC0 | helper cited-only | part of the NRC clade framework referenced in the phylogenetic analysis |
| Rpa1 | sensor comparator | the misclassification produced by the co-scientist parameter set alone |
| Rx | sensor cited-only | named as the activator of the SlNRC3EEE hexamer used as the EM outgroup |
Effectors
No effector is studied. Activation throughout is genetic, via MHD-motif D-to-V substitutions that render helpers autoactive. PVX coat protein is mentioned only in describing the SlNRC3 comparison assembly.
Structures
| Entry | State | Description |
|---|---|---|
| negative-stain EM, not deposited | undecamer (11-mer) | StNRC7EEE+DV ring-shaped particles ~250 A in diameter; 2D class averages resolve 11 peripheral densities |
| negative-stain EM, not deposited | undecamer (11-mer) | SlNRC7EEE+DV and NbNRC7EEE+DV, both 11-mer assemblies, from distinct branches of the NRC7 clade |
| AlphaFold3 predicted | hexamer (modelled) | Hexamer models of 637 non-redundant NRC proteins across 18 clades, three replicates each, with 25 oleic acids as a plasma membrane proxy |
| AlphaFold3 predicted | 5-mer to 11-mer, all low confidence | NRC7 modelled across 5-mer to 11-mer states; no confident assembly at any stoichiometry |
| PDB 9FP6 and PDB 9RI9 (external references) | hexamer | NbNRC2a and SlNRC3 hexameric resistosomes used as the accuracy benchmark |
Open questions
- Whether the NRC7 11-mer functions as a calcium-permeable channel like other characterised CC-NLR resistosomes, or channels other ions or small molecules, is untested.
- Whether the structural expansion of the NRC7 resistosome corresponds to any functional innovation is not addressed.
- There is no high-resolution structure of the 11-mer. The stoichiometry rests on counting peripheral densities in negative-stain 2D class averages.
- NRC4u, the other low-confidence clade, was flagged but never tested. Its models form no ring at all, so what it does is unknown.
- How many of the remaining flagged candidates are genuinely non-canonical is unresolved; only NRC7 was followed up.
- The sensors that activate NRC7 in planta are not identified; activation here is entirely through engineered MHD mutations.
Connections in this corpus
| Paper | Why |
|---|---|
guo-2026-ccg10-resistosome | Two independent breaks from the pentamer/hexamer expectation in the same year, an octamer by cryo-EM and an undecamer by negative-stain EM. Both also propose a way to predict oligomeric state from structure: Guo's HD1-WHD angle heuristic and Toghani's SNI. Read together they argue that resistosome stoichiometry is a variable to be measured, not inherited from the nearest solved structure. |
toghani-2026-sensor-helper-interface | Same first author, same methodological premise, opposite use of model confidence. The sensor-helper paper trusts high-confidence predictions (ipTM 0.759) and validates them by mutagenesis; this one treats low confidence as the informative signal and validates by microscopy. Together they set out when an AlphaFold model should be believed and when its failure is the result. |
seager-2026-avrcap1b-nrc3 | Shared author (Seager) and complementary departures from the canonical hexamer. Seager traps NRC3 below six protomers with an effector; Toghani finds a clade that assembles beyond six natively. Seager's SlNRC3 structure (PDB 9RI9) is one of the two benchmarks against which AlphaFold accuracy is measured here. |
madhuprakash-2026-avrcap1b-tol9a | Shared use of AlphaFold 3 to survey structural variation across a protein family before testing predictions experimentally, Madhuprakash on effector folds and Toghani on resistosome stoichiometry. Both find a clade-level structural distinction that sequence alone did not predict. |
ibrahim-2026-nrg1-chloroplast | Both confront what a low AlphaFold confidence score means. Ibrahim builds a structural argument on an NRG1 model at ipTM 0.43; Toghani argues explicitly that such scores flag candidates worth testing rather than conclusions to draw. The two papers are a useful pair on how far to trust a weak model. |
pai-2026-lettuce-nrc | Both are clade-level analyses of the NRC family built on large-scale phylogenomics, and both find that NRC clades differ in ways that matter functionally. Pai works on evolutionary rate and helper dependency in Asterales, Toghani on structural architecture in Solanaceae. Shared authors Toghani, Kourelis, Contreras, Kamoun. |
marques-2026-legume-tissue | Both scale NLR analysis across many genomes to find structure in a large family, Marques across 28 legume genomes by expression and class, this paper across 346 Solanaceae genomes by predicted assembly. Both depend on consistent NLR annotation pipelines to make cross-genome comparison meaningful. |
gao-2026-vnt1-splicing | Both use AlphaFold3 across a large curated set of NLRs to reason about resistosome assembly, and both treat low-confidence output as informative rather than discardable. Gao reads low pLDDT in N-terminal extensions as evidence they cannot stabilise in the pore; this paper reads low ipTM across a clade as evidence of an atypical architecture. |
Provenance and caveats
The 11-mer stoichiometry comes from counting peripheral densities in negative-stain 2D class averages, not from a high-resolution reconstruction. Negative stain at this resolution supports a protomer count and an overall diameter, nothing finer. No wiki page should describe an NRC7 resistosome structure as solved.
The paper is candid that AlphaFold 3 failed to produce a confident NRC7 model at any stoichiometry tested, including the correct one. The prediction located the anomaly; it did not describe it. This is the paper's central methodological point and should not be softened into 'AlphaFold predicted an 11-mer'.
NRC4u was flagged as equally divergent and never tested. Any claim that SNI predictions were confirmed applies to one of five flagged clades.
This is a v1 preprint, not peer reviewed.
Reference numbers marked [UNVERIFIED] were inferred from in-text context, not read from the reference list.
Sources
- Full citation: Toghani A, Seager BA, Sugihara Y, Roijen L-M, Azcue JM, Garro M, Sargolzaei M, Morianou I, Harant A, Gallop S, Kourelis J, MacLean D, Contreras MP, Kamoun S, Ludke D (2026) AI-guided discovery of atypical protein assemblies. bioRxiv 2026.05.03.722499.
- DOI: 10.64898/2026.05.03.722499
- PDF on file:
papers/2026.05.03.722499v1.full.pdf - Extraction:
data/papers/toghani-2026-sni-undecamer.json(schema v1.0) - References: 31 works, retrieved from OpenAlex
- Licence: CC BY 4.0