When DeepMind’s 2021 AlphaFold paper made protein structure prediction look almost impolitely good, it also exposed a stubborn problem: proteins are not museum statues, they are jittery little shape-shifters with side quests [1]. This new PNAS paper takes that blind spot head-on by asking a very specific, very important question: what if the structure you care about exists only as a rare, fleeting state that normal experiments can sense but not cleanly draw? [2]

The paper focuses on pro-interleukin-18, or pro-IL-18, the precursor to IL-18, an inflammatory signaling protein involved in immune responses [2,3]. Earlier work had already shown that pro-IL-18 is not just sitting there minding its own beta-strands. It samples at least two excited conformations that are very rare, under 0.5% populated, and very short-lived, on the millisecond timescale [2]. In structural biology terms, that is basically the protein whispering from another room.

NMR is one of the few tools that can catch that whisper. It can detect motion across a huge range of timescales and tell you which parts of a protein are involved in the exchange between states [4,5]. But there is a catch, because of course there is a catch. Knowing that certain regions are moving is not the same as having an atomic picture of the moving state. You get clues, not always a full mugshot.

That is the gap this study tries to close.

NMR brings the clues, machine learning brings suspects

Bonin and colleagues use NMR data to narrow down where pro-IL-18 is changing shape, then search through conformational ensembles generated by AlphaFlow, a 2024 machine-learning method built to sample multiple protein conformations rather than one single “best” structure [2,6]. Think of ordinary AlphaFold as the student who turns in one polished final answer. AlphaFlow is the more interesting student who also shows the messy scratch paper.

The key move is that the model does not get to declare victory on vibes alone. The researchers use the experimental NMR constraints to select candidate conformers from the AlphaFlow ensemble, then run additional NMR experiments to see whether those candidates actually fit reality [2]. That last part matters a lot. In biology, “looks plausible” has launched many distinguished careers in being wrong.

The result is a pair of distinct conformers that map onto the two excited states inferred from NMR. In other words, the invisible states became visible enough to model, not because machine learning replaced experiment, but because each method covered the other’s weak side [2].

Why this matters beyond one cytokine

This paper is really about a bigger shift in structural biology. We spent years getting very good at static structures, then remembered that function usually lives in motion. Enzymes open and close. Receptors toggle. Immune proteins mature through transient shapes that may exist just long enough to matter and then disappear like a magician who knows you did not bring a high-speed camera.

Recent work reflects that turn toward dynamics. Reviews in Nature Reviews Chemistry and Current Opinion in Structural Biology argue that ensemble-based, integrative methods are becoming necessary if we want structures that behave like biology instead of like desktop figurines [4,5]. Newer ML systems such as Cfold, AlphaFlow, and BioEmu are all trying to model alternative or equilibrium conformations more directly [6-8].

That is the exciting part. The cautionary part is just as important. Other 2024 work showed that AlphaFold-family models can still miss or mis-handle fold-switched states, and sometimes succeed partly because they have seen similar structures before rather than because they truly understand the underlying energy landscape [9]. The capability gains are real. That is precisely why the validation standard has to stay high. If you are going to trust a model around rare states, hidden intermediates, or possible druggable pockets, “the picture looked convincing” is not a serious safety policy.

The practical upside, with the usual adult supervision

If this hybrid NMR-plus-ML workflow generalizes, it could help researchers map transient protein states involved in signaling, activation, and disease. That matters for drug discovery because rare conformations can expose binding sites that do not show up in the ground-state structure. It also matters for basic immunology, since pro-IL-18 maturation already appears to depend on structural plasticity [3].

Still, this is not a magic microscope. It is one protein, one workflow, and one class of ML-generated ensemble that still depends on training data, model assumptions, and careful experimental filtering. The honest read is not “AI solved protein dynamics.” The honest read is better: AI became useful when it stopped freelancing and started working under lab supervision.

That may be the real lesson here. The future of structural biology probably is not experiment versus machine learning. It is experiment keeping machine learning on a short leash, while machine learning helps experiment look where human intuition would never have bothered to peek.

References

1. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583-589. doi:10.1038/s41586-021-03819-2

2. Bonin JP, Lee JS, Liu ZH, Kim PM, Kay LE. Making invisible excited-state structures of pro-interleukin-18 visible by combining NMR and machine learning. Proc Natl Acad Sci U S A. 2026;123(16). doi:10.1073/pnas.2537014123. PubMed: 41996159

3. Dong Y, Bonin JP, Devant P, et al. Structural transitions enable interleukin-18 maturation and signaling. Immunity. 2024;57(7):1533-1548.e10. doi:10.1016/j.immuni.2024.04.015

4. Wieske LHE, Peintner S, Erdelyi M. Ensemble determination by NMR data deconvolution. Nat Rev Chem. 2023;7:511-524. doi:10.1038/s41570-023-00494-x

5. De Simone A, et al. Integrative approaches for characterizing protein dynamics: NMR, CryoEM, and computer simulations. Curr Opin Struct Biol. 2023;79:102548. doi:10.1016/j.sbi.2023.102548

6. Jing B, Berger B, Jaakkola T. AlphaFold Meets Flow Matching for Generating Protein Ensembles. Proceedings of the 41st International Conference on Machine Learning. 2024. PMLR 235:22277-22303. URL: https://proceedings.mlr.press/v235/jing24a.html and arXiv:2402.04845

7. Buel GR, et al. Structure prediction of alternative protein conformations. Nat Commun. 2024;15:7328. doi:10.1038/s41467-024-51507-2

8. Singh A. BioEmu is a biomolecular emulator for sampling protein structure ensembles. Nat Methods. 2025;22:2008. doi:10.1038/s41592-025-02874-1

9. Chakravarty D, Schafer JW, Chen EA, et al. AlphaFold predictions of fold-switched conformations are driven by structure memorization. Nat Commun. 2024;15:7296. doi:10.1038/s41467-024-51801-z

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.