structuregrade

A mean pLDDT of 72 sounds fine. Two residues stuck at 30 is the part that actually breaks your downstream pipeline.

AlphaFold writes per-residue confidence into the B-factor column, and almost every pipeline collapses it to one mean number before deciding whether to trust the structure. That throws away exactly the information you need: a model can average 72 pLDDT with every residue reasonably confident, or average 72 with a beautifully modeled core and two residues the model was essentially guessing at. Those are not the same structure. structuregrade reads the actual per-residue distribution and reports a grade, a clash count, and the low-confidence tail the mean was hiding.

How it works

Parse PDB ATOM/HETATM records, pull pLDDT straight out of the B-factor field per the AlphaFold convention, bucket residues into confidence bands, and flag anything under 50 by name instead of averaging it away. A separate geometry pass checks inter-residue distances for physically impossible clashes, because a locally confident residue can still be placed somewhere it can't actually be.

What ships

• PDB ATOM/HETATM parser that reads AlphaFold's pLDDT-in-B-factor convention directly
• Confidence banding (>=90 / 70-90 / 50-70 / <50) plus a letter grade, not just a raw mean
• Inter-residue clash detection independent of the confidence score
• CLI, JSON API, browser workbench, Docker, tests

Run it end to end

python -m venv .venv && source .venv/bin/activate
python -m pip install -e .
structuregrade demo
structuregrade serve

Open http://127.0.0.1:8090. Analyze your own JSON input with structuregrade analyze input.json.

API

• GET /api/demo returns the committed fixture and result.
• POST /api/analyze runs the same engine on a JSON body.

The result

The eight-residue fixture averages 71.88 pLDDT — a number that would sail through most "pLDDT > 70" filters. structuregrade assigns it a C instead, because two of those eight residues are below 50 and the mean was quietly burying them. Zero geometry clashes, for what that's worth on a structure this uninterpretable in places.

Update: fixed the clash detector flagging every normal peptide bond

The clash pass compared every pair of atoms from different residues against a 1.5 Å threshold, excluding only atom pairs from the same residue. It never excluded the one bond that legitimately spans a residue boundary: the C(i)–N(i+1) peptide bond connecting adjacent residues along the backbone, which sits at roughly 1.33 Å — well under the "clash" threshold, and completely normal chemistry. On any real, full-atom AlphaFold structure (N/CA/C/O plus sidechains per residue, not the single-CA-per-residue trace the bundled demo happens to use), every adjacent residue pair in the chain would be reported as a clash, silently downgrading grades on structurally perfect proteins.

Verified directly: a two-residue backbone built with standard, textbook bond geometry (including the ~1.33 Å peptide bond) was flagged with 1 "clash" and downgraded from what should be an A to a C, despite nothing being structurally wrong. Fixed by excluding the canonical C(i)–N(i+1) peptide-bond pair between sequentially adjacent residues in the same chain from clash counting, while leaving genuine non-bonded steric clashes between any other atom pair fully detected. tests/test_peptide_bond_clash.py covers the original false positive, confirms a genuine non-bonded clash between unrelated residues is still caught, and confirms the bundled CA-only demo fixture (which never exercised this bond in the first place) is unaffected.

Scope

pLDDT is local per-residue confidence, not interface confidence or domain-placement confidence — PAE and domain-aware review still matter for multi-domain structures and complexes. This is a triage gate that catches "the mean is lying to you," not a replacement for structural validation.

Test

python -m unittest discover -s tests -v

Research basis

• AlphaFold confidence guidance

MIT licensed.