AI & ML interests

None defined yet.

Tc-43 
posted an update 16 days ago
view post
Post
1082
🧬 **An AI agent proposed a novel target, modeled the protein, defined the pocket — then we designed 196 molecules against it.**

**SAMTOR** senses methionine/SAM and gates mTORC1 — the pathway behind methionine restriction, one of the most reproducible lifespan-extending interventions known. No approved drug, no clinical candidate, no published probe.

No human structure has a ligand in the pocket either, so the receptor was threaded onto the 2.10 Å fly holo backbone (7VKR). Pocket Cα RMSD vs human cryo-EM: **1.30 Å**.

196 designs, 2 generations. Ships the receptor *and* the crystallographic cofactor so you can re-dock and disagree with us.

⚠️ Nothing synthesised or assayed. And it's in the card: the generator leaned hardest on TYR249/ARG95 — the least reliable side chains in its own receptor. SAMTOR also *releases* GATOR1 when SAM binds, so filling the pocket might activate mTORC1 instead. No docking score sees that.

🔬 Our viewer Space now serves **13 targets / 5,028 structures** — PDE5, OX2R, SAMTOR, NRAS–CypA and Lp(a) added.

📦 Tc-43/SAMTOR_Novel_Target_Designs_GA-II
🔬 Tc-43/molecular-viewer

CC-BY-4.0 · target proposal + modeling by Claude Code, molecules by TC-43.ai
Tc-43 
posted an update 19 days ago
view post
Post
2309
We are never going to have enough experimental data to train AI for science. That is not a funding problem, it is arithmetic. The PDB has accumulated 258,000 structures in 55 years. One generative run produces more protein–ligand poses than that in a week.

If AI4Science is going to get a data substrate at the scale that made language models work, simulation has to supply it. Experiment will not, at any budget.

But scale is worthless if the labels are wrong, and that is where most synthetic chemistry data quietly fails. A docking score looks like a number you can train on. Often it is not.

Most generated-molecule datasets ship SMILES and nothing else, which makes them impossible to check. I've published 13 that don't.

5,472 de novo small-molecule designs — each with the receptor it was designed against, in the same coordinate frame as the pose, plus an SDF with bond orders and formal charges, per-design properties, and the contact residues for every pose. Load, rescore or re-dock any of them without reconstructing anything.

Targeted protein degradation
· GID4 / CTLH — 1,567
· Cereblon — 221

RAS via cyclophilin A tri-complex
· NRAS·CypA macrocyclic glues (9BG0) — 180
· KRAS·CypA glues — 40

Oncology
· HIF-2α PAS-B — 69
· Cyclin A RxL groove — 57
· WRN helicase, MSI-H synthetic lethal — 34

Cardiovascular & metabolic
· PDE5 catalytic site (2H42) — 2,090
· Lipoprotein(a) / apo(a) KIV-8 (8TCE) — 100
· ACE2·B0AT1 (SLC6A19, PKU) — 38
· Myosin motor domain — 22

Neuro & immune
· OX2R orthosteric, active state (7SQO) — 878
· NLRP3 NACHT — 176

https://huggingface.co/collections/Tc-43/technetium-ga-ii-generativeai-design-sets

cc @hugging-science