Breaking the Pattern with AlphaFold: Introducing the Structural Novelty Index
The AI system AlphaFold has transformed how we predict protein structures. But sometimes the most interesting discoveries come from asking a different question: which proteins break the pattern?
Kudos to AmirAli Toghani, Ben Seager, and Daniel Lüdke, and the rest of the team for introducing the Structural Novelty Index — a new way to use AlphaFold predictions to find protein assemblies that break the pattern. And break the pattern they did: the team uncovered a totally unexpected undecamer, an 11-protomer assembly that is completely different from the canonical hexameric resistosomes of NRC proteins.
That’s not a small tweak. It’s a structural plot twist.
A special shoutout also to our colleagues at Google DeepMind Accelerator for their input and for providing access to co-scientist, which helped sharpen our thinking and expand the SNI framework. AlphaFold pointed us to the unusual assembly, SNI ranked it, co-scientist broadened the search, and electron microscopy showed it was real.
Check the paper:
Toghani, A., Seager, B.A., Sugihara, Y., Roijen, L.-M., Azcue, J.-M., Garro, M., Sargolzaei, M., Morianou, I., Harant, A., Gallop, S., Kourelis, J., MacLean, D., Contreras, M.P., Kamoun, S., and Luedke, D. 2026. AI-guided discovery of atypical protein assemblies. bioRxiv, doi: https://doi.org/10.64898/2026.05.03.722499.
And here is the paper thread originally posted by AmirAli on X/Twitter and Bluesky.
1/n. Protein structure prediction changed biology. Protein assembly discovery is next.
We present SNI (Structural Novelty Index), a scalable framework for finding new protein complexes.
We used it to uncover an unexpected resistosome assembly: 11-mer.
🧵👇🏾
2/n. SNI quantifies AlphaFold models based on features that define canonical protein complexes, incorporating prior knowledge. We applied it to NRC-type NLR immune receptors — typically hexameric — and it cleanly separates canonical hexamers from structural outliers.
3/n. We scaled SNI to 637 non-redundant NRCs from 346 genomes across 85 Solanaceae species.
Most clustered with canonical hexamers. Five clades didn’t.
4/n. Phylogeny and SNI parameters further highlight clades that deviate from canonical hexamers.
SNI flagged NRC7 as the largest, atypical clade in the family. A telltale clue of an unconventional assembly?
5/n. UP TO ELEVEN — We expressed potato NRC7, imaged it by negative-stain EM, ran 2D classification. 🔬💥
Rings ~250 Å across — way bigger than canonical hexamers (~150 Å). 11 protomers.
6/n. One-off, or a clade feature?
We grabbed NRC7 orthologs from tomato and Nicotiana benthamiana. Expressed them, imaged them, counted protomers. Both: 11. It is indeed a feature of the NRC7 clade. 1️⃣1️⃣
7/n. We asked @GoogleDeepMind AI co-scientist to propose SNI parameters based on priors.
It rediscovered 4 of our expert-defined ones, added 3 of its own — and on its own would have flagged NRC7 as atypical. Agentic AI coming to structural biology.
8/n. SNI is a scalable way to discover atypical protein assemblies. Works on any family with prior structural knowledge.
Working on a complex? Give it a try. We’d love to see what you find! Happy to discuss collaborations as well :)
9/n. Huge thanks to an incredible team for making this happen! And to @GoogleDeepMind / @GoogleAI for the co-scientist collaboration. https://www.biorxiv.org/content/10.64898/2026.05.03.722499v1
10/n. One last thing…These go to eleven!
Spinal Tap knew it all along… 1️⃣1️⃣
