CystiNers — Nicholas Foster

3D pocket views and the classifier are coming. This snapshot (built 20 Oct 2026 from the pipeline's feature-vector release) holds every site's descriptors and vectors, but not the protonated structures the 3D cut-out viewer needs or the classifier's results. The quantum vectors and method tabs are complete.

1 · Whole Protein

— res · — atoms

— res >10 Å cut away

2 · 10.0 Å Cut-Out

— res · — atoms

— ≤6Å + — 6–10Å

3 · Peptide Caps

+— HCAP atoms

Heals severed C–N bonds

4 · Buffer Neutralised

— side chains

6–10 Å ions neutralised

5 · GFN2-xTB Water

q(SH)=— · q(S⁻)=—

Target SG + HG (1.002 Å)

STAGE A

Full Folded Protein & 10.0 Å Cut-Out Region

Target SG Inner ≤6 Å Buffer 6–10 Å Cut-Away (>10 Å) Other CYS

STAGE B

Extracted 10.0 Å QM Cluster (GFN2-xTB)

Hover or click any atom or table row below to inspect coordinates, qᵢ, and Fukui fᵢ⁻

Target CYS Inner ≤6 Å Outer 6–10 Å ✂ HCAP ⊕ Neutralised ⋯ H-Bond

1D Sequence & 10 Å Cut-Out Footprint (— aa) —

Target CYS 10 Å Pocket Residue HCAP Cut Sibling CYS (Click)

Click any row to highlight in 3D ·

Residue	Min Dist to SG	Pocket Zone	Charge Handling at pH 7.4	Atoms	Σ qᵢ (Thiol)	Σ fᵢ⁻ (Thiol)	Σ fᵢ⁻ (Thiolate)

Whenever a residue inside the 10.0 Å sphere connects via a peptide bond to a residue outside 10.0 Å, cut_pocket() severs the C–N bond and places a hydrogen link atom (HCAP, pink spheres) along the bond vector (1.09 Å for C–H, 1.01 Å for N–H) for a closed-shell valence structure.

HCAP Atom #	Capped Residue	Backbone Atom	Boundary Severance Type	Bond Len	Dist to SG	3D Coordinates (Å)

⊕ Neutralised Side Chains in 6.0–10.0 Å Buffer

Ionizable residues within ≤ 6.0 Å retain their pH 7.4 charge state from PROPKA. Ionizable side chains between 6.0–10.0 Å (lime green) are neutralised to prevent boundary artifacts.

Residue	Min Dist to SG	Neutralisation Action Applied

⋯ Polar Hydrogen Bonds to Target SG

Hydrogen-bond donors within 3.1 Å of SG stabilize the negatively charged thiolate (-S⁻), lowering $\Delta E_{\text{deprot}}$ and $\text{p}K_a$.

Donor Residue	Donor → H	H···SG Dist	Donor···SG Dist

Cravatt 6-Study Ligandability Matrix:

PER-CYSTEINE 672-DIMENSIONAL FEATURE DECOMPOSITION

Quantum-Chemical Reactivity Profile & 572-D Bag-of-Bonds Spectrum

Active Site

6-Axis Reactivity & Steric Fingerprint

Normalized percentile profile across the computed ligandable proteome

Thiol ($\text{Cys-SH}$) $\rightarrow$ Thiolate ($\text{Cys-S}^-$) Electronic Shift

Direct GFN2-xTB / ALPB(water) comparison before and after proton loss

Observable	Thiol ($\text{-SH}$)	Thiolate ($\text{-S}^-$)	Shift ($\Delta$)

Frontier Orbital Ladder ($\varepsilon_{\text{HOMO}} / \varepsilon_{\text{LUMO}}$) & Sterics

Deprotonation raises $\varepsilon_{\text{HOMO}}$, turning $\text{S}_\gamma^-$ into a potent soft nucleophile

572-Dimensional Bag-of-Bonds (BoB) Pairwise Coulomb Spectrum ($Z_I Z_J / R_{IJ}$)

66 global element-pair bags ($\text{H, C, N, O, S, P, F, Cl, Br, I, Mg}$) $\times$ 8 terms + 11 $\text{S}_\gamma$-anchored local bags $\times$ 4 terms

All 100 Quantum & Electrostatic Pocket Features (34 Direct + 66 Vector Summary)

Every scalar electronic, Fukui, Wiberg, and radial shell descriptor stored for the selected cysteine

Category	Feature Key	Thiol ($\text{-SH}$)	Thiolate ($\text{-S}^-$)	Physical Meaning

2D UMAP MANIFOLD & PROTEIN-GROUPED CATBOOST LIGANDABILITY CLASSIFIER

Quantum + 3D Pocket Chemical Space & Zero-Leakage Evaluation (0 Cysteines)

Zero-Leakage StratifiedGroupKFold

Top-25 / Top-50 Screening Hit Rate — True hit rate on unseen proteins

Low-Score Rule-Out Accuracy — Lowest-scored decile truly non-ligandable

Same-Protein #1 Site Accuracy — Picks true ligandable Cys #1 on a protein

All-Site Binary Accuracy — Unseen test proteins · Confident calls

ROC-AUC vs Single Features — vs SASA, %V_bur & ΔE alone

Experimental Hit Rate Across Score Tiers

Zero protein overlap between training and test sets

Random: 40.5%

Model vs Single-Feature Benchmark

Evaluated strictly on unseen canonical proteins

Model / Predictor	Acc.	Top-50	Same-Prot #1	AUC

Top Physical Drivers (Exact TreeSHAP)

Mean $|\text{SHAP}|$ impact on log-odds across 308 quantum, 3D CIF & sequence features

Top Predictions (Click Row → 3D)

Inspect unseen test hits or cryptic pockets

Site	Split	Truth	P(Lig)	ΔE	%V_bur

18-CORE EXECUTION & SCIENTIFIC ARCHITECTURE

Wave Completion Matrix, Worker Stream & 5-Stage Pipeline

18 Physical Cores · 9 × 2 OMP GFN2-xTB + 10 Boltz-2 Workers

Ligandable Target (9,246) — / 9,246 (—%)

Full mmc3.xlsx Sheet (43,475) — / 43,475 (—%) · Wave —

Boltz-2 + PROPKA Folded Proteins — / 9,714 (—%)

Ordered 100-Row Wave Matrix (Waves 1–100 · Excel Rows 2–10,001)

100% Complete Active Wave Queued

Worker Stream (`run_mmc3_ligandable.py`)

—

Static snapshot: no live worker stream.

STAGE 01100% Sequence Coverage

Sequence & Isoform Resolution

Maps all 9,714 unique proteins (43,475 rows in mmc3.xlsx) across canonical UniProtKB, historical UniSave, and splice isoforms so every residue index verifies against a true C.

STAGE 02Boltz-2 + 700-aa Window

3D Structure Folding

Deduplicates folding at the protein level via Boltz-2. Giant proteins (> 1,400 aa) are folded in a 700-residue window centered on the target cysteine cluster with full index remapping.

STAGE 03PROPKA 3.7 · pH 7.4

pH 7.4 Protonation & H-Bonds

PDB2PQR + PROPKA assigns physiological pH 7.4 protonation states and optimizes the 3D hydrogen-bond network while filtering detached terminal H artifacts (> 1.45 Å).

STAGE 0410.0 Å Sphere + 6.0 Å Buffer

10 Å Pocket Cut-Out & Capping

Extracts all residues within 10.0 Å of $\text{S}_\gamma$, caps severed peptide bonds with HCAP hydrogens, reconstructs missing $\text{H}_\gamma$ protons on low-$\text{p}K_a$ thiols, and neutralises side chains beyond 6.0 Å.

STAGE 05GFN2-xTB / ALPB(water)

Dual-State QM & 672-D Vector

Runs GFN2-xTB single-points + condensed Fukui ($f^+, f^-, f^0$) on both Thiol ($\text{-SH}$, charge $q$) and Thiolate ($\text{-S}^-$, charge $q-1$) plus Morfeus SASA & buried volume.