Tc-43/GPR75_Intracellular_Antagonist_Designs_GA-II

GPR75 Intracellular Antagonist Designs — Technetium GA-II

156 small molecules generated de novo by the Technetium `TC-43.ai` engine (GA-II), conditioned on the intracellular allosteric site of GPR75 — a class A orphan GPCR and first-in-class anti-obesity target with human genetic validation. Each molecule was constructed against this pocket rather than selected from a compound library — docking (AutoDock Vina) came afterwards, to place and score the generated molecules in the site. They are built as antagonists / inverse agonists that block Gα_q coupling. Each design is supplied as a full protein–ligand complex, docked into an inactive-state receptor model, with the SMILES, physicochemical properties and site-engagement metrics used to triage it.

This is the companion set to `Tc-43/GPR75_Inhibitor_Designs_GA-II`, which targets the extracellular vestibule. The two sets address different pockets on the same receptor and are not interchangeable.

Provenance

1. Why the intracellular site — the design rationale

1.1 The therapeutic direction is fixed, and it is inhibition

GPR75's validation is human genetics. In ~640,000 exomes, protein-truncating variants occur in about 4 per 10,000 people, and heterozygous carriers are leaner and obesity-protected (BMI ≈ 0.34 SD lower, ~5.3 kg lower body weight, obesity OR ≈ 0.46). Because loss of function is protective, the rational pharmacology is inhibition or inverse agonism. GPR75 also shows modest constitutive Gα_i activity, so suppressing basal tone — not merely blocking an agonist — is the objective.

1.2 The orthosteric pocket is not available

The orthosteric site is occluded by ECL2 and does not open:

A drug-like ligand needs roughly 1.7 Å. Solvent-accessible orthosteric volume was 0 ų in all 10 MD frames. The channel does not open in 150 ps of restrained dynamics. Designing into the orthosteric pocket is not an option for this receptor.

1.3 The intracellular allosteric pocket is a validated site class

The canonical class A intracellular allosteric pocket — bounded by TM1, TM2, TM3(3.50), TM6, TM7(NPxxY) and helix 8 — has direct chemical precedent:

Ligands here block G-protein engagement rather than competing with an orthosteric agonist — precisely the mechanism required for a constitutively active receptor with no accessible orthosteric site.

1.4 Why the inactive state specifically

The pocket is present in all three receptors but is best formed in the inactive conformation, where TM6 packs back in and the Gα groove closes into a discrete cavity:

(Canonical reference set: the 18 Ballesteros–Weinstein positions lining the CCR2 / CCR9 / CXCR2 / C5aR1 / β₂AR intracellular site.)

There is a mechanistic reason as well as a geometric one. An inverse agonist works by stabilising the inactive conformation. Docking into an active, G-protein-coupled structure would select for compounds that fit the state the program is trying to suppress. The inactive model is therefore the appropriate design receptor, and 9XQC is used here only to measure Gα_q occlusion, which requires the α5 helix to be present.

1.5 How the inactive receptor was built, and how far it can be trusted

Consensus per-helix active→inactive transition derived from five class A structure pairs (β₂AR 4LDE/6PS2, A_2A 5WF5/5NM4, 5-HT_2A 8UWL/7WC8, M₂R 7T94/5ZKC, NTSR1 8FMZ/6YVR), transferred onto 9XQC via GPCRdb generic numbering, then restrained minimisation and 150 ps MD.

Validation: leave-one-out RMSD reduction of 15–38% across the five pairs, transfer scale k\ = 0.88 ± 0.19. The transfer captures roughly 25% of the true active→inactive difference.* This is a modelled conformation, not an experimental structure, and it is the principal uncertainty in this dataset. See §5.

1.6 The pocket was found independently, not only by anchored search

Two detection methods were run:

1. 1.PSP enclosure grid (0.7 Å spacing, 1.4 Å probe, 7-direction protein–solvent–protein count, enclosure ≥ 5/7), anchored on the NPxxY/H8 landmarks Y376 (7.53), A381 (8.48), D80 (2.40).
2. 2.Blind Voronoi alpha-sphere clustering (fpocket-style, whole protein, no anchors, radius 3.2–6.5 Å, single-linkage 3.5 Å).

The blind method ranks this pocket #1 of 15 in both the experimental 9XQC and the inactive model. Cavity centroids agree to 0.3 Å between the two receptor states, with 72% shared lining. The cavity stays open in all 10 inactive-ensemble snapshots (48–75% of cavity points remain >3.0 Å from protein), so it is not a single-snapshot artefact.

1.7 The mechanistic link to Gα_q

In the deposited 9XQC complex, the Gα_q α5 helix C-terminus penetrates this volume — N354, L355, R356, E357, Y358, N359, L360, V361, with a closest approach of 0.07 Å (Y358) and 32 Gα atoms within 2.5 Å of the cavity. A ligand occupying this pocket is directly competitive with Gα_q engagement.

Every design in this set was checked against that helix after superposing the design receptor onto 9XQC. All 156 sterically overlap the α5 C-terminus (median closest approach 0.65 Å, minimum 0.26 Å), occluding 4.8–14.9% of the α5 envelope volume. Both figures are reported per-design in designs.csv.

1.8 The site definition used for docking

The full cavity is a broad groove; the docking site is the TM1/TM2/TM7/H8 lobe — 149.2 ų, 58% apolar, 26 lining residues.

Lining: G59(1.49) I62(1.52) V63(1.53) F64(1.54) L65 S66(1.56) F67(1.57) D69(1.59) A71 F72 R76 T77 F79(2.39) D80(2.40) I83(2.43) L84(2.44) S87(2.47) R143(3.50) T322(6.36) P373(7.50) Y376(7.53) S377(7.54) R378(7.55) N379 S380 A381(8.48)

Required anchor set — the only three contacts made by all ten reference chemotypes in the pharmacophore panel: I83 (2.43), Y376 (7.53), R378 (7.55), with D80 (2.40) as the polar floor. All 156 designs engage all four. Median lining coverage is 15 of 26 residues.

2. What is in this dataset

designs.csv columns

design_id, vina_kcal_mol, ligand_efficiency, smiles, mw, clogp, tpsa, hbd, hba, rotatable_bonds, heavy_atoms, rings, aromatic_rings, fsp3, qed, formal_charge, lipinski_violations, anchor_I83, anchor_Y376, anchor_R378, polar_floor_D80, pocket_lining_contacts_of_26, gaq_a5_volume_occluded_pct, gaq_a5_closest_approach_A, murcko_scaffold, contact_residues, ga_run, structure_file

SMILES are as emitted by the generator (REMARK SMILES in each complex PDB). The SDF poses were rebuilt by assigning bond orders from those SMILES onto the docked coordinates; all 156 round-trip identically to the source SMILES with stereochemistry removed, and 98 match with stereochemistry included — the remainder are 3D poses that define centres the input SMILES left unspecified.

3. Chemistry of the set

107 distinct Murcko scaffolds across 156 designs. Formal charge distribution: 118 neutral, 22 anionic, 16 cationic. Ten independent GA runs contributed; two account for 57% of the output.

4. Suggested use

Rank by gaq_a5_volume_occluded_pct rather than by vina_kcal_mol if the objective is Gq blockade — the two are only loosely related, and the docking score is a ranking device, not a binding free energy. pocket_lining_contacts_of_26 and the four anchor flags give an orthogonal, score-free measure of site engagement.

Cross-docking against receptor_GPR75_inactive_ensemble.pdb is the recommended robustness check — it tests whether a compound's anchor set survives the receptor model's own conformational uncertainty.

4a. Use as synthetic data for AI/ML

This is a machine-generated dataset: every molecule was produced by a generative engine and every label was computed, not measured. That makes it usable for a few things and unsuitable for others.

Reasonable uses

• —Pretraining or fine-tuning pocket-conditioned generative models — 156 ligands paired with a single fixed receptor and a defined site, each with a full complex PDB.
• —Property regression from structure — designs.csv carries MW, cLogP, TPSA, HBD/HBA, rotatable bonds, rings, aromatic rings, Fsp³, QED and formal charge alongside SMILES.
• —Pose-based learning — 156 3D ligand conformations in a common receptor frame, with per-design contact residue lists and geometric site-engagement metrics.
• —Scaffold-diversity work — 107 distinct Murcko scaffolds over 156 molecules.

What this is not

• —Not affinity data. vina_kcal_mol is an empirical docking score used here for ranking. It is not a measured binding free energy, and training a model to predict it teaches the model the scoring function, not the biology.
• —Not a benchmark. There is no held-out split, no defined task, and no experimental labels to score against.
• —Not conformationally validated. Poses come from rigid-receptor docking; none has been refined with dynamics or rescored by an independent method.

The geometric columns — pocket_lining_contacts_of_26, the four anchor flags, and gaq_a5_volume_occluded_pct — are measured directly from coordinates rather than predicted by a scoring function, so they are the most defensible labels in the file.

5. Limitations — read before using

• —Nothing here has been synthesised or assayed. These are computational designs.
• —There is no experimental inactive-state GPR75 structure. The docking receptor is a model that captures ~25% of the true active→inactive transition. This is the largest single uncertainty in the dataset.
• —The lobe boundary is a choice, not a measurement. The rank-1 alpha-sphere cluster is large and fused; restricting the site to the TM1/TM2/TM7/H8 lobe at 9 Å from the centroid was a judgement call.
• —The scoring function has no electrostatic term. Whether an anionic head group is required — as it is for several known class A intracellular antagonists — cannot be decided from these scores. 22 of the 156 designs are anionic; that is an untested hypothesis, not a result.
• —Structure cannot distinguish inverse agonism from neutral antagonism. Occupying this pocket predicts blockade of Gα_q coupling. Whether that reads out as suppression below basal requires a functional assay.
• —The "arginine basket" is effectively one arginine. R378 (7.55) is pocket-facing and engaged by all 156 designs. R76 is buried (9.6 Ų side-chain SASA, guanidinium hydrogen-bonded into three TM2 backbone carbonyls); R143 (3.50) and R384 (8.51) are solvent-exposed but point away from the cavity, with only 19% and 6% of their neighbouring space reachable from the ligand volume. Do not design toward R76, R143 or R384 in this receptor.

Citation

Technetium Therapeutics (TC-43.ai). GPR75 Intracellular Antagonist Designs — Technetium GA-II.
Hugging Face, 2026. CC-BY-4.0.
Target rationale: Apodex AI. Receptor and pocket modelling: Claude Code (Anthropic).
Generative design: Technetium TC-43.ai engine.

Released under CC-BY-4.0.