Skip to content

Assignment

Backbone resonance assignment

Assemble spin systems from triple-resonance spectra and place them on the sequence, with calibrated per-residue confidence.

Assignment is the bottleneck in most structural and dynamics projects, and it is the step where a confident mistake is most expensive: everything downstream inherits it, and nothing downstream detects it.

Our approach is probabilistic throughout. Peaks are grouped into spin systems, typed by their carbon shifts, and placed against the sequence by a search over the whole assignment rather than by greedy sequential walking — so a locally attractive but globally impossible placement gets rejected. What comes back is not a single answer but a distribution, and the per-residue confidences are calibrated, meaning a residue reported at 90% is right about 90% of the time.

Because the confidences are real, the software can tell you what to do about the low ones. Cartouche ranks the experiments that would most reduce the remaining ambiguity, so the next week of spectrometer time is spent on the residues that need it rather than on a uniform re-run.

Cartouche is benchmarked on public data, not on a flattering subset: across 87 proteins and 8,046 residues it places backbone assignments correctly 89% of the time pooled (91% median), in the same range as the best published deep-learning assignment method on the same proteins. Where it can, it also draws on a sequence-only secondary-structure prior from a protein language model (ESM-2), which sharpens placement when peak lists are sparse.

What it needs

  • Triple-resonance peak lists (HNCACB, HNCO, HNCA and relatives)
  • Protein sequence

What you get

  • Assigned chemical shifts
  • Per-residue confidence
  • A ranked list of experiments to acquire next

Enter wherever your data already is

This analysis runs as a full pipeline from spectra, or picks up from work you have already done — so nothing has to be repeated to get to the next step.

  • From spectra
  • From a peak list
Enter a pipeline wherever your data already isA four-stage pipeline runs left to right: spectra, peak picking, peak fitting, rate fitting. Three labelled entry points feed in from below — full starts at spectra, from-picking starts at peak fitting, and from-fitting starts at rate fitting — so work already done elsewhere does not have to be repeated.SpectraPeak pickingPeak fittingRate fittingfullfrom-pickingfrom-fitting

What does the work

  • Cartouche

    Built in-house

    Our probabilistic assignment with calibrated confidence

  • SSC

    Built in-house

    Constraint-solver spin-system clustering