Can AI modelling of protein structures distinguish between sensor and helper NLR immune receptors?
AmirAli ToghaniiD, Raoul FrijtersiD, Tolga O. BozkurtiD, Ryohei TerauchiiD, Sophien KamouniD, Yu SugiharaiD
Abstract
NLR immune receptors can be functionally organized in genetically linked sensor-helper pairs. However, methods to categorize paired NLRs remain limited, primarily relying on the presence of non-canonical domains in some sensor NLRs. Here, we propose that the AI system AlphaFold 3 can classify paired NLR proteins into sensor or helper categories based on predicted structural characteristics. Helper NLRs showed higher AlphaFold 3 confidence scores than sensors when modelled in oligomeric configurations. Furthermore, funnel-shaped structures--essential for activating immune responses--were reliably predicted in helpers but not in sensors. Applying this method to uncharacterized NLR pairs from rice, we found that AlphaFold 3 can differentiate between putative sensors and helpers even when both proteins lack non-canonical domain annotations. These findings suggest that AlphaFold 3 offers a new approach to categorize NLRs and enhances our understanding of the functional configurations in plant immune systems, even in the absence of non-canonical domain annotations.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 15
- FWCI
- 7.00
- Field percentile
- top 2%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (29 Nov 2024) · CC BY 4.0
AlphaFold 3 oligomer models separate paired plant NLRs: helpers score higher pTM and form funnel-shaped resistosomes unlike sensors, including some rice pairs lacking integrated-domain annotations.
Computational methods can currently tell sensor and helper NLRs in genetically linked pairs apart only when the sensor carries a non-canonical integrated domain (ID), which only some sensors have and annotation tools can miss. The work tests whether AlphaFold 3 (AF3) oligomer confidence and resistosome funnel prediction offer a structure-based alternative that does not depend on IDs.
Key results
ShownAcross eight benchmark pairs whose sensors carry an integrated domain, helpers received consistently higher AF3 predicted template modelling (pTM) confidence than their paired sensors in both pentameric and hexameric oligomer models.
Benchmark: rice Pikm, Pii, Pia; orthologs PIK5/6-NP, Pi5-3/1, Pias; barley RPG5/HvRGA1, RGH2/3. N-terminus to NB-ARC end modelled as 5x and 6x with 50 oleic acids and three seeds; helper pTM exceeded sensor pTM in both stoichiometries (values in Table S2; ipTM also compared, Figs. S3-S4). · Helper NLRs produce higher AlphaFold 3 pTM…; Figure legends
SuggestedHelper oligomer models consistently contained the funnel-shaped structure thought to insert into membranes and drive immunity, whereas sensors assembled only low-confidence resistosome-like structures lacking a funnel.
Visual inspection of seed-1 pentamer and hexamer models (ChimeraX) for all eight benchmark pairs (Fig. 1c; Figs. S5, S6). No quantitative funnel metric or per-seed structural comparison is reported. · Helper AF3 structures form funnel-shaped…; Figure legends
ShownMADA motif screening, the sequence-based alternative, classified only the Pikm and PIK5/6-NP pairs, whose helpers exceeded the HMM cut-off. All other benchmark sensors and helpers lacked the motif, leaving those pairs unassigned.
MADA HMM run with HMMER v3.4 (--max); motif called at HMM score > 10 (Table S3). Authors conclude AF3-based classification is more reliable than this sequence-based approach, since AF3 pTM separated helpers from sensors across the benchmark. · Helper AF3 structures form funnel-shaped…; NLR annotation; section 14
ShownIn five rice Nipponbare CC-NLR pairs where one member has an ID annotation, the ID-free partner (putative helper) scored higher pTM and formed a funnel, while the ID-carrying putative sensor did not, matching the benchmark pattern.
Ten head-to-head linked, cross-clade rice pairs with full CC domains, mapped to RefSeq GCF_034140825.1 via DIAMOND BLASTP; AF3 pentamers and hexamers with 50 oleic acids, three seeds (Fig. 2b,c; Figs. S7-S9; Table S7). · AF3 can classify paired NLRs even when the…; Curation of NLR sequences
SuggestedFor the five ID-free rice pairs, roles were assigned by averaged pentamer/hexamer pTM. This placed PIK5-NP (Pair-13) as sensor despite its unannotated HMA domain, and Pair-07 showed a consistent putative sensor-helper pTM difference.
InterProScan misses the integrated HMA domain of PIK5-NP, yet pTM ranking agreed with its known sensor role. Of the remaining four ID-free pairs, Pair-07 is the one reported with a consistent pTM gap (Fig. S7; Table S7). The rice assignments are not experimentally tested. · AF3 can classify paired NLRs even when the…; NLR annotation
Methods, data and reproducibility
Coiled-coil (CC) NLR sequences from the N-terminus through NB-ARC were modelled on the AlphaFold 3 webserver as pentamers and hexamers with 50 oleic acids (membrane proxy) over three seeds. Within each pair, pTM/ipTM scores and funnel presence were compared, contrasted with MADA motif HMM screening, then applied to uncharacterized rice NLR pairs.
- Benchmark CC-NLR pairs with ID-carrying sensors: rice Pikm, Pii, Pia, their orthologs PIK5/6-NP, Pi5-3/1, Pias, and barley RPG5/HvRGA1 and RGH2/3 · rice (Oryza sativa) and barley · n = eight NLR pairs
- Rice cv. Nipponbare protein set (Oryza_sativa_vg_japonica.protein.fasta) and a prior list of head-to-head linked, cross-clade NLR pairs (Supplementary Data 6 of a cited study) · Oryza sativa cv. Nipponbare · doi:10.7946/P2FC9Z
- Matched Nipponbare RefSeq NLR pairs retained for AF3 after requiring a full N-terminal CC domain (five with and five without an ID annotation) · Oryza sativa cv. Nipponbare · n = ten CC-NLR pairs · GCF_034140825.1
- NLRtracker-annotated NLRs from RefSeq proteomes of two dicots and 11 monocots plus RefPlantNLR, used for an NB-ARC phylogeny of monocot paired NLRs · multiple plant species (two dicots, 11 monocots) · n = 4,936 NLR proteins from 13 proteomes; 5,509 NB-ARC sequences
- AF3 input sequences and models plus NLRtracker and HMMER outputs generated in this study · 10.5281/zenodo.13826775
- Methods & tools
- AlphaFold 3 (webserver), NLRtracker v1.0.3, InterProScan v5.67-99.0, HMMER v3.4 with MADA motif HMM, DIAMOND BLASTP v2.1.9, RefPlantNLR, refplantnlR, ChimeraX, SciPy (Pearson correlation)
- Software & environment
- AlphaFold 3 webserver (golgi.sandbox.google.com); NLRtracker v1.0.3; InterProScan v5.67-99.0; HMMER v3.4 (--max); DIAMOND BLASTP v2.1.9; ChimeraX and SciPy used without stated versions.
- Reproducibility gaps
- Models came from the hosted AlphaFold 3 webserver, whose model version is not stated. Released scripts cover only the phylogenetic analysis, and per-model pTM/ipTM values sit in supplementary tables not included in the text.
Limitations (3)
- InterProScan missed the integrated HMA domain of PIK5-NP, so its domain architecture had to be manually reassigned from CNL to CONL; automated ID annotation alone was insufficient for sensor labelling. · stated by the authors
- Classification relies on relative pTM ranking with no reported decision threshold or statistical test. Among ID-free rice pairs other than Pair-13, only Pair-07 is described with a consistent gap, and none of the rice assignments are validated experimentally. · noted in this summary
- The benchmark is small (eight pairs), models are truncated at the NB-ARC end, funnels are judged visually on seed-1 models, and no singleton NLR controls are included. Rice pairs lacking a full N-terminal CC domain (Pair-02, Pair-03, Pair-08, Pair-09, Pair-10) were not modelled, so such pairs remain untested. · noted in this summary
Who should read:Read this if you work on paired plant NLRs or rice/barley resistance genes and want a structure-based way to label sensors and helpers with AlphaFold 3 when integrated-domain annotations are missing.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- What NLR you recognizing? Predicted binding affinities- and energies may be used to identify novel NLR-effector interactions
Alicia Fick, Jacobus Lukas Marthinus Fick, … Noëlani van den Berg·The Plant JournalPeer-reviewed·Apr 1, 2025
- A conserved structural logic underlies sensor-helper NLR communication in the NRC immune receptor network
AmirAli Toghani, Maián Garro, … Mauricio P. Contreras·bioRxiv·May 20, 2026
- The Pik NLR pair accumulates at the plasma membrane as a hetero-oligomeric sensor-helper immune protein complex prior to activation
Hsuan Pai, Mauricio P. Contreras, … Clémence Marchal·bioRxiv·Dec 2, 2025
- Computational design of de novo integrated domains enables rational control of pathogen effector recognition in plant NLR immune receptors.
Yuxuan Xi, Angus Bucknell, … Adam R. Bentham·bioRxiv·Jul 21, 2026
- Structural basis for the broad recognition specificity of an Arabidopsis immune receptor
He Zhao, Camille-Madeleine Szymansky, … Jonathan D. G. Jones·bioRxiv·Feb 7, 2026
- Functional divergence shaped the network architecture of plant immune receptors
C.-Y. Huang, Y.-S. Huang, … C.-H. Wu·bioRxiv·Dec 12, 2023