Skip to content

Analysis: paracetamol

Five notebooks over the runs the cMD and REST2 pages produced.

notebook what it shows
pmf-1d.ipynb the free energy along each of the three torsions
clustering.ipynb symmetry-first clustering: the symmetry enumerated, the distance built from it, HDBSCAN, the 18-of-20 assignment, populations and representative structures
clustering-comparison.ipynb the historical cluster-then-merge workflow, the new vote alone, and symmetry-first, on the same frames
mutual-information.ipynb which torsions move together, and the estimator's own bias
symmetry-alignment.ipynb the same fit, every frame written in the labelling nearest the density peak of the largest cluster: and the ring pair

Committed with their outputs, and rendered here as pages: the figures and numbers come from the runs these tutorials describe. To run them yourself:

micromamba activate analysis-env          # see Installing, step 5
export MD_TUTORIAL_RUNS=/path/to/your/tutorial/runs
jupyter lab docs/tutorial/paracetamol/analysis/

Why paracetamol is the interesting case

Three torsions, and the REST2 convention treats them differently. The amide omega and all six aromatic ring bonds are protected — unscaled in every state — so what REST2 heats are the bonds OUTSIDE the ring: aryl (the ordinary torsion about the exocyclic C(aryl)–N bond, which is not part of the ring) and phenol (the hydroxyl rotation).

The phenyl ring also has a two-fold axis, which makes this the smallest honest example of a symmetry-aware distance (methods):

  • the ring flip RENAMES the two sides of the ring, so a frame measured through C5 and the same frame measured through C10 are one configuration; the distance is the minimum over those namings, and symmetry-related frames are at distance zero before HDBSCAN fits anything;
  • the flip is the coupled case: it moves aryl and phenol together, and both of its images (1-3-4-10 and 9-7-8-17) fall outside the torsion table, so they are recomputed from coordinates -- not approximated by a 180 degree shift, which on this trajectory would be wrong by up to 25 degrees;
  • the fit finds two clusters, A at mass 0.514 and B at 0.484, with 0.225 % density noise and nothing ambiguous. The four namings of two conformers never become four clusters.

The previous page, archived unchanged, found four clusters and merged them into the same two groups afterwards; the comparison notebook shows the two workflows agree on 3998 of 4000 frames and why the other two differ.