[us-commits] [ehb54/ultrascan3] a8bd96: Add perceive validate: predict the coded residues ...

[emre brookes]

  Branch: refs/heads/ehb54-issue-980
  Home:   https://github.com/ehb54/ultrascan3
  Commit: a8bd967b8319fd9605340a9b23db8126f1820abf
      https://github.com/ehb54/ultrascan3/commit/a8bd967b8319fd9605340a9b23db8126f1820abf
  Author: ehb54 <brookes at uthscsa.edu>
  Date:   2026-08-08 (Sat, 08 Aug 2026)

  Changed paths:
    M us_somo/develop/include/us_hydrodyn_perceive_somo.h
    M us_somo/develop/perceiver/DECISIONS.md
    M us_somo/develop/src/us_hydrodyn_perceive_somo.cpp
    M us_somo/develop/src/us_hydrodyn_script.cpp

  Log Message:
  -----------
  Add perceive validate: predict the coded residues as if unknown

Rebuilds every residue somo.residue does code, pretending it is unknown, and
reports computed vbar, molvol and hydration against the stored values with a
summary. The same idea as perceive compare, which does this for atom typing,
but for the computed properties, and it exercises the whole pipeline
including perception. That makes it the only end-to-end measure of accuracy.
Chain termini and residues with unmodelled atoms are counted and skipped
rather than quietly averaged in.

Over seven structures and 1157 rebuilt residue instances, ordinary amino
acids give mean absolute errors of about 2.7% on vbar and 2.1% on molvol,
matching what the standalone harness claimed.

The ligand rows are the interesting ones. Heme's computed volume is 32.6%
above the stored value and NAD's is 28.1% above. Earlier work in this branch
reached the same conclusion by a completely different route, using 3V both on
the isolated molecule and by difference in context, and arrived at comparable
numbers. Two independent methods agree the stored ligand volumes are around
30% low. NAD's computed vbar of 0.619 is also closer to the measured 0.620
than the stored 0.599 is.

Heme's vbar comes out 24% low, which is the expected failure: it is a
metalloporphyrin, Durchschlag and Zipper publish no increment for iron, and
the engine reports the atom for review rather than inventing a value.

Hydration is much the weakest of the three, around half the residues within
half a water and systematically low, the misses being the polar and charged
side chains. A per-type majority cannot express residue-specific hydration.
This also confirms the pH coupling from the runtime side: residue_list
carries pH-adjusted hydration, so aspartate reads 6.0 and glutamate 7.0,
exactly the charged-carboxylate values, where the raw file fields give 1.0.

The lookup counts one vote per residue name, since residue_list holds several
entries for residues with ionization variants and counting each weights them
more heavily purely for being pH-aware, which was enough to flip a borderline
type and cost serine and threonine a water each.

Refs ehb54/ultrascan-tickets#980

Co-Authored-By: Claude Opus 4.8 <noreply at anthropic.com>



To unsubscribe from these emails, change your notification settings at https://github.com/ehb54/ultrascan3/settings/notifications