n-mourdou/ADMETrix — reverse-engineered prompt

Build me a working local version of this ADMETrix project. I want it to generate new small molecules that are optimized across multiple ADMET properties, and also let me do scaffold hopping from a seed SMILES so I can try to reduce toxicity while keeping similarity to a starting compound.

Please wire up the full flow so the generator talks to the ADMET prediction service in real time, calculates rewards, runs the optimization loop, and logs experiments and outputs in MLflow. I’d like it to be easy to start, ideally with the services coming up through Docker and the main experiment scripts ready to run from the repo with clear commands. If a GPU is available, use it, otherwise make sure CPU mode still works.

Also make sure there is a simple way to swap in my own seed molecule, run both the ADMET optimization and scaffold hopping examples, and verify the API and tracking UI are actually running. If anything is unclear, look up the current docs for the linked tools and make sensible fixes so it runs cleanly.