Post
2279
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
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