Fig 4A
Geometric scanning recall
Sγ–Sγ geometric scanning of AFDB-monomer models recovers only a small fraction of disulfides annotated in matched PDB experimental structures, even at generous distance cutoffs.
Paper Fig 4
AlphaFold geometric scanning
AlphaFold-monomer geometric scanning is fundamentally limited as a disulfide-bond predictor. AFDB-monomer models do not place Sγ atoms within the geometric tolerance of an experimental disulfide bond often enough to serve as an annotation source: well-conserved disulfides in PDB experimental structures are repeatedly placed beyond the Sγ–Sγ cutoff in matched AFDB models, even when sequence identity to a templated PDB structure is high.
Population statistics
Aggregate evidence across paired AFDB / PDB structures.
Fig 4B
Sγ–Sγ distance distributions
Distribution of Sγ–Sγ distances in AFDB models for cysteine pairs that form an experimental disulfide in the matched PDB structure. The bulk of the distribution sits well above the 2.5 Å disulfide bond cutoff.
Fig 4C
PAE attenuation
Predicted aligned error between disulfide-paired cysteines does not predict whether the AFDB model places them within geometric tolerance, i.e., low PAE alone is not sufficient to recover the bond geometry.
Structural examples
Three case studies where a disulfide is firmly geometric in the PDB experimental structure but absent from the matched AFDB-monomer model. Download the PyMOL sessions to inspect the cysteine-pair geometry locally.
Fig 4D
Example 1 — AFDB vs PDB
Cartoon overlay of an AFDB-monomer model and matched PDB experimental structure showing a disulfide that is correctly geometric in the PDB structure but absent from AFDB at any reasonable cutoff.
Fig 4D · structure assets
Paper-supplementary identifiers pending.
Fig 4E
Example 2 — AFDB vs PDB
Second AFDB / PDB pair with the same failure mode: an experimentally observed disulfide is placed beyond Sγ–Sγ geometric tolerance in the AFDB model.
Fig 4E · structure assets
Paper-supplementary identifiers pending.
Fig 4F
Example 3 — AFDB vs PDB
Third AFDB / PDB pair illustrating the limitation across structurally distinct domains.
Fig 4F · structure assets
Paper-supplementary identifiers pending.
Methods note
Sγ–Sγ geometric scanning uses a 2.5 Å cutoff between cysteine sulfurs to call a disulfide. Scanning AFDB-monomer models with the same cutoff recovers a small minority of PDB-annotated disulfides; relaxing the cutoff to 5 Å trades sensitivity for a high false-positive rate. The point of this page is not that AlphaFold is wrong — the residue-level conformations are reasonable — but that monomer geometry is the wrong inference layer for disulfide annotation, which is why TriCyp uses ESM2-3state predictions over geometric scanning for AFDB-source domains.