Explore molecules invented by a GPT. The network groups similar ones, and each can be seen in 2D and in 3D.
The GPT writes the molecule one piece at a time, in SMILES format. You pick the logP and QED bins and the model writes to match the request. Each piece shows a bar with the confidence the model had in that choice. Hover over a piece to see the other options it considered. The writing was recorded from the model and only the animation speed is artificial.
Comparison between 4000 molecules generated by the GPT and a real ZINC250k sample, measured with RDKit.
| Measure | GPT | Real ZINC |
|---|
When a logP bin is requested, the mean logP of the valid molecules the model returned.
| Requested bin | Mean logP obtained | Valid |
|---|
The loss measures how much the model misses the next piece. Both lines falling together mean it learned without memorizing.
| Step | Train | Validation |
|---|