Tc-43/SAMTOR_Novel_Target_Designs_GA-II

SAMTOR — SAM-competitive de novo designs (Technetium GA-II)

196 small molecules across two generations, generated de novo by the Technetium TC-43.ai engine (GA-II) and conditioned on the S-adenosylmethionine (SAM) pocket of human SAMTOR. 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. With no approved drug, clinical candidate or published chemical probe for this target, there was no library to screen — every molecule here was built against the pocket. Released as docked protein–ligand complexes with interaction profiles, physicochemical descriptors and predicted ADME.

SAMTOR is a novel, previously undrugged target. It was proposed, researched and structurally modeled by Claude Code, working from public deposits — the target hypothesis, the human holo receptor model, the pocket definition and the design brief all originate from that analysis rather than from an existing programme. The molecules were generated by the Technetium TC-43.ai engine (GA-II), then docked, profiled and triaged.

To our knowledge this is the first public set of small-molecule designs against SAMTOR. There is no approved drug, no disclosed clinical candidate and no published chemical probe for this target.

Two generations, one anchor

	batch 1	batch 2
designs	100	96 new + 4 carried elites
central heterocycle	7-azaindole (100/100)	quinoline (96/96 new)
spiro-cyclopropane isoindolinone	100/100	100/100
Vina median	−12.9	−14.0
ligand efficiency	0.40	0.42
H-bonds per design	3.14	4.31
PHE203 H-bond	97%	100%

The two generations share one element and differ in the other. The spiro[cyclopropane-1,1'-isoindolin]-3'-one is retained in 196/196 designs — it is the group that satisfies the PHE203 backbone NH, the single most reliable anchor in the pocket. The central heterocycle it hangs off was swapped wholesale between generations, azaindole → quinoline. Batch 2 scores ~1.1 kcal/mol better and the gain is polar, not lipophilic (corr(vina, clogp) = +0.06).

Use the batch column (1, 2, 1+2) and central_heterocycle to slice the two series. The four 1+2 rows are byte-identical elites carried forward by the GA.

Why SAMTOR

SAMTOR is the direct molecular sensor coupling cellular methionine/SAM status to mTORC1 activity. When SAM is abundant it binds SAMTOR, SAMTOR releases the GATOR1–KICSTOR complex, and mTORC1 signals for growth; when SAM is scarce, GATOR1 stays active and mTORC1 is suppressed. Dietary methionine restriction is among the most reproducible lifespan-extending interventions across species, and a SAM-competitive SAMTOR ligand is a route to phenocopy it pharmacologically — one layer upstream of mTOR, on a single nutrient-input branch, rather than the indiscriminate suppression a rapalog delivers.

The receptor

No human SAMTOR structure with a ligand in the pocket exists. The only human structure (9O5D, cryo-EM 3.34 Å, within the KICSTOR–GATOR1 assembly) is apo; every liganded structure is Drosophila. The receptor here is a human holo model built by threading the human sequence (UniProt Q1RMZ1) onto the 2.10 Å fly holo backbone of 7VKR chain A, so the ligand-bound backbone conformation is experimental rather than predicted. It is identical across all 196 complexes.

Property	Value
Modeled residues	216 (human numbering 93–393, gapped)
Sequence at modeled positions	100% human
Backbone source	7VKR chain A, 2.10 Å (holo)
Threading identity over the core	35.6% (77/216 identical)
SAM-contact shell identity, fly → human	11 / 14
Pocket Cα RMSD vs human cryo-EM (9O5D)	1.30 Å (1.55 Å under a core-domain fit with no refitting)
Steric clashes	0
Polar contacts to SAM ≤ 3.5 Å	14

The pocket and the pharmacophore

Orthosteric shell (human numbering): R95 · G172 · S173 · C174 · D190 · I191 · V192 · C201 · D202 · F203 · S244 · L245 · L246 · Y249. Adenine, ribose and the methionine arm are each read out by a separate residue set — the tripartite signature of a cofactor site, giving three semi-independent anchors.

Across all 196 poses:

Residue	Interaction	Frequency
PHE203 backbone NH	H-bond to the isoindolinone lactam	193 / 196
TYR249	face-to-face π-stack, 3.6–3.9 Å centroid	133 / 196
ARG95	cation–π	112 / 196
ILE191 / LEU246	hydrophobic floor	196 / 196

91 designs engage all three polar anchors at once. All 196 fully occupy the adenine and ribose sub-sites, with ligand centroids a median 0.8 Å from SAM's, and median buried SASA of 75.9% against 82.4% for SAM itself — near-cofactor burial with more heavy atoms.

Files

File	Contents
designs.csv	196 rows × 51 columns
ligands.sdf	196 docked poses, 3D, bond orders and formal charges verified against the generation SMILES
receptor_SAMTOR_human_model.pdb	the shared rigid receptor
reference_SAM_cofactor.pdb	crystallographic SAM, for competitiveness analysis
structures/	196 complex PDBs (receptor + LIG, with REMARK SMILES and CONECT)

Column groups

• Identity — design_id, rank, batch, central_heterocycle, smiles (as generated, stereo undefined), isomeric_smiles_pose (configuration as docked), murcko_scaffold, pose_file
• Docking — vina_score (kcal/mol), ligand_efficiency, buried_sasa_pct, sa_score (Ertl synthetic accessibility)
• Interactions — total_interactions, hbonds, hydrophobic, pi_stacking, cation_pi, halogen_bonds, n_contact_residues, contact_residues (;-separated), plus anchor_phe203_hbond, anchor_tyr249_pistack, anchor_arg95_cationpi
• SAM competitiveness — sam_centroid_dist_ang, sam_adenine_occupancy, sam_ribose_occupancy, sam_methionine_arm_occupancy
• Physicochemical — mw, clogp, tpsa, qed, hbd, hba, rotatable_bonds, heavy_atoms, aromatic_rings, fsp3, formal_charge
• Predicted ADME — hlm_clint_ml_min_g, hepatocyte_clint, logd_7_4, logs_aqueous, half_life_h, ppb_pct, caco2_logpapp, herg_prob, cyp3a4_inhib_prob, cyp2d6_inhib_prob, dili_prob, ames_prob, bbb_prob

Usage

from datasets import load_dataset
ds = load_dataset("Tc-43/SAMTOR_Novel_Target_Designs_GA-II", split="train")

import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_csv(hf_hub_download("Tc-43/SAMTOR_Novel_Target_Designs_GA-II", "designs.csv", repo_type="dataset"))
df.groupby("central_heterocycle")[["vina_score", "hbonds", "fsp3", "logs_aqueous"]].median()

# designs standing on the reliable anchor, with tolerable developability
hits = df[(df.anchor_phe203_hbond == 1) & (df.hlm_clint_ml_min_g < 30) & (df.logs_aqueous > -5)]

from rdkit import Chem
mols = [m for m in Chem.SDMolSupplier("ligands.sdf") if m]   # 196 3D docked poses

Limitations — please read before using

• Entirely computational. Nothing here has been synthesised or assayed. These are design hypotheses.
• Rigid-receptor docking. Vina scores rank poses; they are not binding free energies.
• The receptor is a model, and its reliability is uneven across the pocket. Backbone-mediated contacts reproduce the human cryo-EM structure well (F203 1.05 Å, S244 0.43 Å side-chain deviation); key side chains do not — R95 2.49 Å, D202 2.33 Å, Y249 4.76 Å. 9O5D is 3.34 Å, so those rotamers are poorly determined there too; this is uncertainty rather than proven error. But contacts to those three residues cannot be validated against anything.
• Batch 2's improvement rests disproportionately on that uncertain group. Its extra H-bonds are side-chain, not backbone: side-chain H-bonds per design rose 1.35 → 2.65 (R95 52→93, D202 21→73) while backbone H-bonds fell slightly, 1.79 → 1.66. The backbone fraction dropped from 57% to 39%, and designs standing on ≥2 backbone H-bond residues fell from 52 to 41. Rank on backbone-mediated contacts if you want the ordering that survives a better receptor.
• Developability drifted the wrong way between generations. Fsp³ 0.24 → 0.19, cLogP 3.92 → 4.58, aromatic rings 4.0 → 4.5; logS < −5 in 56 → 65 designs, hERG > 0.5 in 91 → 98. Across the full set, 119/196 have logS < −5 and 185/196 have hERG > 0.5. CYP3A4 improved (0.82 → 0.65). Raising Fsp³ remains the lever that would move solubility, hERG and CYP together.
• ADME predictions are out of applicability domain. ADMET-AI is trained on approved drugs; these are MW ≈ 440 tri- and tetra-aromatics. Treat the values as a ranking, not as predictions.
• A mechanistic caveat specific to this target. SAMTOR's function is that SAM binding releases GATOR1. A ligand that occupies the pocket and stabilises the SAM-bound conformation would release GATOR1 and activate mTORC1 — the opposite of the therapeutic intent. Occupancy alone is not the design goal, and nothing here has been tested against that criterion. 191 of 196 designs fully occupy the methionine-arm sub-site; only 5 cover less than 60% of it, so the set barely explores the occupy-without-triggering hypothesis. Every docking score and interaction count in this file would look equally good for a compound that does the opposite of what is intended.
• Target validation for SAMTOR is open. Whether mTORC1 suppression via this branch alone reproduces methionine-restriction phenotypes is unestablished.

Provenance

• Target proposal, literature research, structural modeling and pocket definition — Claude Code (Anthropic), from public RCSB/UniProt/AlphaFold deposits.
• Molecule generation — Technetium TC-43.ai engine, GA-II generation (two batches).
• Docking — AutoDock Vina, rigid receptor. Interaction profiling — Tc-PLIP. Descriptors — RDKit. ADME — ADMET-AI.

Source structures: 7VKR, 7VKQ, 7YRJ, 9O5D, AlphaFold AF-Q1RMZ1-F1. Human sequence UniProt Q1RMZ1.

Citation

@misc{technetium_samtor_gaii_2026,
  title  = {SAMTOR SAM-competitive de novo designs (Technetium GA-II)},
  author = {Technetium Therapeutics},
  year   = {2026},
  note   = {Target proposed and modeled by Claude Code; molecules generated with the TC-43.ai engine},
  url    = {https://huggingface.co/datasets/Tc-43/SAMTOR_Novel_Target_Designs_GA-II}
}

Released under CC-BY-4.0. Part of the Technetium GA-II GenerativeAI Design Sets collection.