A missing measurement is usually treated as a nuisance. In protein nuclear magnetic resonance, it can also be evidence. When part of a protein moves between conformations on microsecond-to-millisecond timescales, its signal can broaden until it is too weak to assign. Dyna-1 was trained to recognise which residues are likely to be missing for that reason.[1]
The researchers began with assignment records from roughly 10,000 proteins in the Biological Magnetic Resonance Data Bank and curated more than 100 independent NMR relaxation datasets. The striking result was that models trained on missing assignments also predicted exchange measured by those separate relaxation experiments.[1]
The evidence behind Dyna-1
| Measure | Value | Context |
|---|---|---|
| Large weak-label source | About 10,000 deposited proteins | Missing chemical-shift assignments provide a large but imperfect training signal.Source 1 |
| Curated evaluation data | 133 R1/R2/NOE datasets plus 10 CPMG datasets | The public repository identifies these as RelaxDB and RelaxDB-CPMG.Source 2 |
| Model input | Sequence and/or structure | The best public version uses an intermediate representation from ESM-3.Source 2 |
| Output | Per-residue motion probability | It estimates microsecond-to-millisecond exchange rather than reconstructing every conformation.Source 2 |
Why a moving residue can disappear
NMR observes nuclei in their chemical environments. If a residue stays in one environment long enough, it contributes a comparatively sharp signal. If it switches between environments at an awkward rate, the signal is spread and weakened. The peak may become difficult to detect even though the residue is physically present.
That creates a useful inversion. Instead of asking only what the visible peaks say, the model asks whether the pattern of absences looks like exchange broadening. The assumption is not always true—signals can be missing for technical and experimental reasons—so the value comes from testing the learned pattern against direct relaxation measurements.[1]
Function leaves a dynamic signature
Dyna-1 performed especially well on motion connected to enzyme catalysis and ligand binding. The researchers also found that residues undergoing this form of exchange tended to be more evolutionarily conserved. That is biologically plausible: motion can be part of the mechanism, not merely thermal noise.[1]
What the model can and cannot answer
| Question | Dyna-1 contribution | Remaining evidence needed |
|---|---|---|
| Which residues may move on μs–ms timescales? | A probability for each residue. | Targeted relaxation or other experiments for confirmation.Source 2 |
| Does motion matter to function? | Prioritisation based on learned patterns and conservation. | Mechanistic experiments linking the motion to catalysis or binding.Source 1 |
| What exact conformations occur? | Not supplied by the output. | Structural and kinetic measurements or simulation.Source 1 |
| Can one sequence variant change dynamics? | The model can compare sequences on a shared backbone. | Experimental validation of the variant’s physical behaviour.Source 2 |
The larger lesson: weak labels can hide in scientific archives
The project did not create a vast new motion experiment from scratch. It recognised that an old database already contained a proxy signal in what researchers had failed to assign. That is a powerful pattern for scientific AI: find a large, imperfect observation, train on it, then test whether it predicts a smaller set of expensive, direct measurements. The proxy becomes valuable only when that second step succeeds.
Sources & further reading
- Learning millisecond protein dynamics from what is missing in NMR spectra Nature · 10 Aug 2026 · Accessed 16 Oct 2026
- Dyna-1: Model for predicting micro-millisecond motions from protein sequence and/or structure GitHub · Accessed 16 Oct 2026