3 ms·
It appears that the molecules on the page are generated with a machine learning algorithm trained on a small organic molecules dataset that is heavily biased to
by honie 5y ago
It appears that the molecules on the page are generated with a machine learning algorithm trained on a small organic molecules dataset that is heavily biased towards cyclic, particular nitrogen heterocyclic, structures (I only sampled about 40 of them, so it could also have just been my luck).
The model seems to have learnt chemistry pretty well because, as many have already pointed out, most of the molecules generated do actually exist (or are extremely likely accessible if they haven't already been documented). Even the ones with strange bond angles have otherwise perfectly normal number of bonds. The only time where I get molecules that cannot possibly exist are those with overlapping atoms that just defy known physics.
Addendum: it is worth noting that the model might actually have been trained with data that contain bond lengths, or even spatial information if no post-generation geometry optimisation is performed before a molecule is rendered.
- messe 5y agoSo https://thischemicalprobablyexists.com https://thischemicalprobablyexists.com might be a better title? I'm not a chemist, I've an undergrad degree in physics/mathematics. Intuitively, your answer sounds right, but I'm not in a position to judge for sure.
- bee_rider 5y agoChemistry is just the integral of physics (he said, extremely sophomoricly), so you should be able to work it out.
- rflrob 5y agoThis chemical could probably be made but doesn’t exist seems most accurate. https://www.chemistryworld.com/opinion/chemical-space-is-big-really-big/7899.article https://www.chemistryworld.com/opinion/chemical-space-is-big...
- CamperBob2 5y agoI'll wait for https://iwontworkwiththischemical.com https://iwontworkwiththischemical.com.
- whymauri 5y agoFor demos like this, it's pretty common to just run RDKit for a structural check before serving the user the actual chemical. I don't know what kind of model this is, though. They could have limited SMILES (popular choice) to have a more stable generative space or they might have introduced validity into the loss. I think the coolest part is how the builder got all the rendering to work well! Or it could also be a rule-based fragment model guaranteed to hit a valid structure, that works too.
- isoprophlex 5y agoI did exactly this once: trained a simple language model (Karpathy's Char-RNN iirc) on a txt file with SMILES strings. After validating the output, it was easy to plot a (2D) skeletal formula. I never got around 3D renders, but I guess a SMILES -> 2D -> 3D pipeline with some molecular mechanics structure energy minimization for the 3D part is cheap to do. I found the output surprisingly diverse. Model was very good in adding branched lipid tails that kept rambling on forever, though...
- ta988 5y agoThose overlapping atoms are just artifacts of the energy minimization they used that was not able to find a proper conformation. They may have filtered molecules with highly defavorable energies to avoid that.