# ESMFold2 Disulfide-Axis Reviewer-Proofing

No new ESMFold2 inference was run in this pass; analyses use saved target-level outputs.

## Axis

High disulfide capacity is the top 20% of exact-sequence primary rows: `disulfide_capacity_per100 >= 1.3193`.
Total rows: 652; high-disulfide rows: 134.

## 1. Cluster-Aware Validation

| k-mer Jaccard cluster threshold | clusters | largest cluster | one-rep effect | one-rep CI | cluster-mean effect | cluster-mean CI | cluster-bootstrap row CI |
|---:|---:|---:|---:|---|---:|---|---|
| 0.50 | 518 | 10 | -0.117 | [-0.166, -0.069] | -0.117 | [-0.166, -0.069] | [-0.218, -0.079] |
| 0.35 | 489 | 19 | -0.118 | [-0.169, -0.070] | -0.118 | [-0.167, -0.071] | [-0.218, -0.081] |
| 0.25 | 477 | 21 | -0.121 | [-0.172, -0.071] | -0.121 | [-0.172, -0.072] | [-0.219, -0.082] |

This is approximate k-mer clustering, not a replacement for final MMseqs2/CD-HIT clustering, but it directly tests the family-overrepresentation objection.

## 2. Matched Controls

High-disulfide rows were greedily matched 1:1 to lower-disulfide controls on source family and viral/nonviral status, then nearest log length, TM/hydrophobic proxy.

| slice | pairs | mean high-control CA-lDDT | CI | median | high worse fraction | exact source+viral pairs |
|---|---:|---:|---|---:|---:|---:|
| all_matched | 127 | -0.061 | [-0.107, -0.016] | -0.034 | 0.638 | 125 |
| cameo_or_cameo_followup | 12 | -0.066 | [-0.222, 0.131] | -0.139 | 0.750 | 12 |
| external_rcsb | 17 | -0.103 | [-0.245, 0.040] | -0.037 | 0.706 | 17 |
| fresh_rcsb_general | 42 | -0.098 | [-0.159, -0.041] | -0.028 | 0.690 | 40 |
| viral_rcsb | 56 | -0.020 | [-0.090, 0.054] | -0.015 | 0.554 | 56 |

## 3. Confidence/Calibration

| axis | n | top n | top CA-lDDT | lower CA-lDDT | top failure rate | lower failure rate | top pTM | lower pTM | pTM AUC good | top overconfident failure rate |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| high_disulfide | 652 | 134 | 0.713 | 0.862 | 0.306 | 0.083 | 0.638 | 0.770 | 0.967 | 0.000 |
| classII_E_E2_envelopeE | 289 | 11 | 0.396 | 0.803 | 0.818 | 0.165 | 0.287 | 0.710 | 0.953 | 0.000 |

## 4. Per-Source Visualization

Figures:

- `figures/high_disulfide_by_source.png`
- `figures/high_disulfide_quartile_curve.png`
- `figures/ptm_vs_ca_lddt_high_disulfide.png`

## 5. Structure Examples

The final scaled evaluations did not save predicted coordinates, so three selected failures were rerun with coordinate saving under the same inference-time scaling setting: `biohub/ESMFold2`, 10 loops, 68 sampling steps, 5 diffusion samples, seed 0, best-by-pTM selection. The outputs are CA-only superposed reference/prediction overlays, not full-atom publication renderings.

| target | best sample | CA-lDDT | pTM | CA-RMSD | overlay PNG | overlay PDB |
|---|---:|---:|---:|---:|---|---|
| 9B6W_1_A | 0 | 0.228 | 0.203 | 34.08 | `example_structures/predicted_best_by_ptm/9B6W_1_A.ca_overlay.png` | `example_structures/predicted_best_by_ptm/9B6W_1_A.ca_overlay.pdb` |
| 8R2L_1_AAA | 3 | 0.238 | 0.185 | 31.37 | `example_structures/predicted_best_by_ptm/8R2L_1_AAA.ca_overlay.png` | `example_structures/predicted_best_by_ptm/8R2L_1_AAA.ca_overlay.pdb` |
| 8R8L_1_A | 4 | 0.252 | 0.179 | 28.87 | `example_structures/predicted_best_by_ptm/8R8L_1_A.ca_overlay.png` | `example_structures/predicted_best_by_ptm/8R8L_1_A.ca_overlay.pdb` |

## Artifact Index

- `analysis_outputs/cluster_aware_disulfide_effects.csv`
- `analysis_outputs/cluster_assignments.csv`
- `analysis_outputs/matched_control_summary.csv`
- `analysis_outputs/matched_control_pairs.csv`
- `analysis_outputs/confidence_calibration_summary.csv`
- `analysis_outputs/ptm_calibration_bins.csv`
- `analysis_outputs/per_source_high_disulfide_summary.csv`
- `example_structures/example_structure_targets.csv`
- `example_structures/example_prediction_overlay_summary.csv`
- `example_structures/predicted_best_by_ptm/*.ca_overlay.png`
- `example_structures/predicted_best_by_ptm/*.ca_overlay.pdb`