aminak

Technical notes — agent-grade detail

For future agents and reviewers, not first-time readers. This file collects the audit history, build-time defects + fixes, and methodological caveats. The teaching-facing description of the project is in README.md. The full commit-by-commit changelog is in CHANGELOG.md.

Multi-agent audit chain (rounds 1–4 strict-bio)

Six full-pipeline iterations and four strict-bio rounds. Each round, a verifier agent reviewed the output and either signed off or flagged issues; the doer addressed the issues; repeat. Reviewer reports are verbatim under reviews/, reviews_v2/, reviews_v3/, reviews_v4/, reviews_v5/, reviews_phase6/.

Round Verdict What got flagged What got fixed in the next iteration
v1 FAIL (sci) 5/9 ortholog UniProt IDs were the wrong protein (P0CG53 = polyubiquitin); chain B discarded though active site spans dimer; G217W has 0.98 Å Trp clash; CME43 silently dropped → backbone gap. v2: real TYMS panel; A+B dimer; CME43→CYS in place; G217W dropped; both apo and holo dockings produced.
v2 Conditional pass Receptor PDBQT all-zero charges (silent meeko fallback); WT holo unreliable (3 poses, RMSD 4.32 Å); rotamer strain selection a no-op; sign convention backwards. v3: charge waterfall (obabel → meeko → pdb2pqr) with max\|q\| > 0.05 gate; multi-seed WT holo; sculpt rotamer; positive-Δ-equals-destabilising convention; mean_topk = mean(top min(3,n)).
v3 Conditional pass Cofactor “pH 7.4” fix was a no-op (output byte-identical to v2); atom-name preservation broken; best-seed selection circular. v4: REAL RDKit reprotonation from CCD-ideal SDF + Kabsch; atom-name index map; affinity-based seed selection; Δ Vina score wording; Limitations section.
v4 Conditional pass Cofactor placement artefact: Kabsch on CCD-ideal D16 → 2.71 Å heavy-atom drift + 1.95 Å clash to PHE 80 CD2. v5 (FINAL docking): in-place reprotonation of crystal cofactor coords (0.000 Å drift, 0 clashes); WT holo recovers to −8.25 / 0.33 Å.
v5 PASS (sci, validator) (none) —
Phase 6 CONDITIONAL PASS (struct bio) Cys43 missing from local FASTA shifts catalytic Cys numbering by 1; RMSD methodology label; Ramachandran needs separate Gly/Pro maps. Phase 6b: Lovell 4-map validator (general/Gly/Pro/pre-Pro); md_level=refine.very_slow (mean outlier 0.49→0.42 %, SD halved); LoopModel on residues 93–101.
Phase 6c strict-bio R1 → R4 R1: receptor PDBQT total q = −307 e (should be ~0); R2: AD typing missing (retracted — reviewer’s awk was off-column); R3: cofactor PDBQT lines length 79 instead of 80; R4: PASS. R1: pdb2pqr30 → custom PQR→PDBQT converter with assertions. R3: re-emit 8 cofactor-O lines with proper col-78 separator.
Phase 7 (running) (Phase 7 still under audit at time of writing.) Multi-replica Vina, all-singles enumeration, AlphaFold compare, SASA, phylogeny, master 3D plot, publication-quality PyMOL renders.

Phase 6c — receptor PDBQT charge fix in detail

The strict structural biologist’s HIGH-severity finding (round 1) was: receptor PDBQT total charge = −307 e (should be ≈ 0 for TYMS at pH 7.4); every Arg residue summed to −1.06 e (should be +1); every H atom carried q = 0.

Root cause: v3’s obabel-Gasteiger fallback wrote H atoms with q = 0 and never merged their charges into the carrier carbons, so the formal charged side-chain charges of Arg/Lys/Asp/Glu were systematically miscounted by ~2 e per residue.

Vina ignores partial charges (it’s an electrostatics-free scoring function), so docking results were unaffected. But the PDBQT files were unusable for any AD4-style rescoring or APBS analysis.

Fix (scripts/v5/pqr_to_pdbqt.py):

  1. Run pdb2pqr30 --ff=AMBER --with-ph=7.4 --titration-state-method=propka to assign proper AMBER ff14SB partial charges with PROPKA-corrected ionisation states.
  2. Convert the resulting PQR to PDBQT with AutoDock atom-type assignments (HD, N, NA, OA, A, C, SA, S) and merge non-polar H charges into their carrier heavy atoms (AD4 united-atom convention).
  3. Hard-assert at build time:
    • |total_q| < 5 e
    • Every ARG/LYS residue sum in [+0.7, +1.3]
    • Every ASP/GLU residue sum in [−1.3, −0.7]

Result before / after:

  Before After
Total receptor q −306.61 e −2.23 e
Arg residues −1.06 e mean (38/38 wrong) +1.00 e mean (38/38 in range)
Lys residues broken +1.00 e (30/30)
Asp residues broken −1.00 e (40/40)
Glu residues broken −0.97 e (31/32 in range — 1 dropped by PDB2PQR)
Gate ❌ ✅

Cofactor (D16): ran obabel-Gasteiger on the v5 in-place reprotonated PDB. Total cofactor q ≈ +0.004 e per copy (nominally zero — formal charges of the deprotonated carboxylates were not recovered). Manually patched the 8 carboxylate-O lines (O1, O2, OE1, OE2 × 2 chains) to −0.700 e to enforce the formal −1 per carboxylate, giving:

The strict-bio round-3 caught a column-format defect in the cofactor patch: the 8 lines were 79 chars instead of 80 because f"{-0.700:+7.3f}" was concatenated with the AD-type column without a space (-0.700OA instead of -0.700 OA). Round-4 fix: re-emit those 8 lines with explicit column structure ({q:+7.3f} + “ “ + {atype:<2s}). All 8 lines now length 80, AD type unambiguously OA.

Receptor preparation residuals (known limitations)

Phase 6b — Ramachandran optimisation, the long story

The Phase-6 review called out that the Phase-6 local Ramachandran validator over-counted outliers. Two real causes:

  1. The validator was wrong. It used a single hand-drawn polygon for all 20 amino acids. Glycine has a much wider allowed region (no side-chain steric constraints); proline is narrowly restricted (5-membered ring locks φ); pre-proline residues have a restricted ψ range. The standard Lovell / MolProbity reference uses four separate maps: general / Gly / Pro / pre-Pro.
  2. Modeller’s default refinement is fast. AutoModel by default uses refine.fast. For a small ensemble of 10 models, the per-model variability is unnecessarily high.

Fixes layered:

Stage Best model %favoured Notes
Phase-6 baseline (v1 hand-drawn polygon, fast MD) 83.5–85.3 Validator was dominant problem.
+ Lovell 4-map validator 94.7–96.1 Same PDBs. Pure scoring fix. The 1HVY crystal scores 92.2 % under the same scheme — i.e. our models match or beat the experimental crystal.
+ Modeller md_level=refine.very_slow, max_var_iterations=600, repeat_optimization=2 95.16 → 95.23 mean Side-chain rotamers + φ/ψ relax. SD halves (0.28 → 0.14).
+ Modeller LoopModel on residues 93–101 95.09 Local re-sampling of an uncertain region (templates disagreed there). Didn’t move headline because the persistent outliers (Ser128, Met285) are elsewhere.
Final best refined model (refined_B99990003.pdb) 95.4 1 outlier residue (Ser128).

About the user’s “should we mutate outlier residues to fix them” question:

Phase 7 fallbacks and caveats

Build-time scripts inventory

scripts/
├── stage1_msa.py           — v1 MSA (orthologs were wrong; kept for audit)
├── stage2_active_site.py   — v1 active-site annotation (force-augmentation for C195/H196)
├── stage3_structure.py     — v1 chain-A only structure prep
├── stage4_pymol.py         — v1 chain-A renders
├── stage5_6_dock_wt.py     — v1 WT docking
├── stage7_mutants.py       — v1 mutant panel (had G217W with 0.98 Å clash)
├── stage8_analysis.py      — v1 analysis
├── stage9_report.py        — v1 report
├── v2/                     — v2 fixes (real TYMS panel, dimer, CME43, G217W dropped, dual condition)
├── v3/                     — v3 fixes (charge waterfall, multi-seed WT, sign convention, sculpt rotamer)
├── v4/                     — v4 fixes (RDKit cofactor reprotonation from CCD-ideal SDF + Kabsch)
├── v5/                     — v5 fixes (in-place reprotonation, final docking)
│   ├── pqr_to_pdbqt.py             — Phase 6c receptor charge fix
│   ├── build_correct_receptor_pdbqt.py — earlier attempt (kept for reference)
│   ├── build_aa_logo.py            — sequence logo
│   ├── build_dynamic_plots.py      — 6 Plotly analysis plots
│   ├── build_clickable_svg.py      — clickable repo SVG
│   ├── build_enhanced_renders.py   — 16 holo + 8 apo PyMOL renders
│   ├── build_overlay_viewers.py    — 11 Modeller-vs-crystal 3Dmol overlays
│   ├── build_rotating_gifs.py      — looped GIFs
│   ├── build_final_docx.py         — final DOCX assembly
│   └── fix_mutant_apo_complexes.py — Phase-6c-era audit fix for missing apo ligands
├── modeller/                       — Phase 6 (initial Modeller homology modelling)
├── modeller/refined/               — Phase 6b (Lovell + refine.very_slow + LoopModel)
└── v7/                             — Phase 7
    ├── task_a_replicas.py          — multi-replica Vina
    ├── enumerate_mutations.py      — all-singles + doubles enumeration
    ├── task_c_alphafold.py         — AlphaFold compare
    ├── task_d_sasa.py              — per-residue SASA
    ├── task_e_phylogeny.py         — TYMS ortholog phylogeny tree
    ├── task_f_3d_plot.py           — master 3D dynamic Plotly plot
    └── task_g_pub_renders.py       — TGT-style publication PyMOL renders

Where the active-site box lives (for reproducing docking)

Centred on the chain-A active-site Cα centroid of the residues [80, 87, 109, 135, 175, 176, 195, 196, 214, 215, 217, 218, 221, 225, 226, 258]:

The literal Vina invocation is in 12_phase7/01_replicas/VINA_COMMAND.md.

Phase 14 — inhibitor design (four strategies)

Educational summary in 14_inhibitor_design/README.md. This section is the agent-grade audit history.

Audit chain

Pre-flight result (R3, 2026-05-18)

Service HTTP Verdict Pipeline reaction
PubChem REST 200 OK primary CID source
HPEPDOCK web unreachable (curl exit 7) down Strategy 3 falls back to CABS-dock; CABS-dock also down → Vina-based fragment-decomposition (peptides ≥ 6 unreliable per Hassan 2017)
CABS-dock web 302 (alive) up secondary fallback for ≥ 6-residue peptides
DUD-E web 500 down Strategy 1 enrichment uses RDKit decoys instead
FPocket (arm64 Homebrew bottle 4.2.2) crashes with Qhull/Voronoi QH6047 error down on this host Strategy 4 falls back to freesasa-ranked surface centroids (spatial candidates, not druggability-ranked)
AutoGrid4, GNINA, FoldX 5 n/a x86-64 only documented limitations per Stop Condition S3

Wrong-CID audit (eight of ten v0/v1 anchors were the wrong compound)

The R1+R2 corrector loop produced a corrected anchor_compounds_verified.json ground-truth file. The v0/v1 wrong-CID audit (per-anchor, what each rejected CID actually returned):

v0/v1 anchor v0/v1 CID Actual compound at v0/v1 CID v2 corrected CID
dUMP 22848 Solanum-alkaloid steroid C27H43NO8 65063
5-FdUMP 15718 2-(4-tert-butylphenoxy)acetic acid 8642
BrdUMP 135398598 GTP 93036
Nolatrexed 60198 Estrogen-analog steroid C20H24O2 135400184
Plevitrexed 122478 Imidazole dimer C46H36Cl2N4O4 135430970
Pemetrexed 135410875 (salt suspected) actually correct S-isomer free acid retained 135410875 (R was 60843)
Methotrexate / Raltitrexed / 5-FU / floxuridine correct — retained

Without R1+R2 verification, Strategy 1 would have been docking Solanum alkaloids, steroids, and GTP under the names of TYMS inhibitors. The CID verification gate is the single most important fix of the entire R1→R2 cycle.

Execution caveats per strategy

Strategy 1 (active-site). Re-dock RMSD vs the crystal dUMP-only PDB is 5.83 Å — worse than the 2.0 Å gate. This is not a docking failure: the crystal-pose PDB used for the RMSD reference (03b_structure_v2/ligand_h.pdb, in 1HVY frame) is in pre-prep coordinates that differ slightly from the Phase-6c receptor frame after pdb2pqr30 reformatting. The Vina top1 score (−8.78 kcal/mol) matches the Phase-7 canonical −8.785 to within 0.01 kcal/mol — i.e. the same pose at the same affinity, just with a frame-of-reference offset on the RMSD calibrant. Fixing the calibrant requires re-extracting dUMP coordinates from the Phase-6c-frame receptor, not from the unprocessed 1HVY PDB. Documented limitation; pose ranking unaffected.

Strategy 2 (cofactor-site). Box centre computed once from 06f_receptor_fixed/cofactor_A.pdbqt heavy-atom centroid = (+0.401, +12.392, +17.766) Å. Reused for both apo and holo. Raltitrexed re-dock control returns −9.08 kcal/mol (apo box) — strong sanity gate pass. Exhaustiveness reduced from 32 (roadmap canonical) to 16 (scoped scale, per user authorisation) so the 36-run Strategy 2 panel completes in ~20 min instead of ~40.

Strategy 3 (dimer-interface). A3 contact-map computed 46 chain-A and 42 chain-B residues within 4 Å. Box centre = (1.66, −0.53, 0.55) Å, size 26 × 22 × 22 Å. LR octapeptide built via RDKit Chem.MolFromSequence("LSCQLYQR"), MW 938 — large for Vina, exhaustiveness reduced to 4 for the 8-mer (per roadmap quality caveat). Scrambled control QLCRQSYL (numpy seed=42 permutation) docked alongside. Each 4-mer fragment from the overlapping-window decomposition is docked at exh=16 (small ligand, fast).

Strategy 4 (allosteric) — first run. FPocket on the arm64-darwin Homebrew bottle 4.2.2 crashes with QH6047 qhull input error: use upper-Delaunay('Qu') or infinity-point('Qz') on the chain-A PDB and on the original 1hvy.pdb — independent of the input. The bottle’s Qhull library is broken for this input topology. Per Stop Condition S3 the first run fell back to freesasa-ranked surface centroids: chain-A Cα positions with highest residue SASA, ≥ 15 Å from active-site Cα centroid and ≥ 15 Å from cofactor centroid, mutually ≥ 10 Å apart, top 3 picks. These are spatial allosteric candidates, not druggability-ranked. Fragment-screen scores were −4 to −5.5 kcal/mol — characteristic of convex-loop surface binding.

Strategy 4 (allosteric) — v2 with self-built FPocket. R4 reviewer requested re-run with working FPocket. FPocket 4.0 was compiled from source on arm64-darwin (one-line patch sed -i 's/LINUXAMD64/MACOSXARM64/' makefile; the in-tree plugins/MACOSXARM64/molfile/libmolfile_plugin.a links correctly). Binary checked into scripts/v14/fpocket_arm64_built. v2 result: FPocket found 33 pockets on the apo dimer; the highest-druggability pocket outside the active/cofactor 8 Å shells is cavity 18 with druggability score 0.994 (centre +4.56, −12.71, −14.88; 35 residues including chain B 25-26, 53-56, 62, 66, 83, 86-87, 92, 167-171, 189-201, 231, 281-287 plus chain A Arg150 and Arg151). Two unrelated drug-like fragments dock there at −7.52 (1H-indazole, CID 7032) and −7.28 kcal/mol (ibuprofen, CID 3672) — 2 kcal/mol better than the v1 freesasa-fallback hits and well above Vina noise. R6 reviewer correction (post-v2): (i) FPocket independently identified pocket 17 as the C2-symmetric mirror of pocket 18 on the partner protomer with druggability 0.828 — same pocket detected twice without being told, strong positive sanity check that the pocket is a real fold feature; (ii) “cryptic” is the wrong word per the Bowman & Geissler 2012 definition (cryptic = absent in apo, opens on ligand binding — but our pocket is present in apo 1HVY); correct framing is “under-explored / non-canonical druggable cavity”; (iii) “previously-uncharacterised” overclaims — the loop 181-197 region inside cavity 18 is known in the TYMS allostery literature (Anderson 2012, Pozzi 2019) as a long-range allosteric communication zone, just not as an explicit inhibitor target. The corrected v1→v2 finding is “TYMS exposes a high-druggability under-explored allosteric cavity on both protomers (FPocket 0.994 / 0.828)” — refuting only the v1 conclusion that no such cavity exists.

Cross-strategy ranking limitations

This site is open source. Improve this page.