swansonk14/SyntheMol — reverse-engineered prompt

Build me a Python tool for drug discovery that can generate small molecule candidates that are not just predicted to be active, but also realistic to make in the lab. I want it to explore a big reaction based chemical space using purchasable building blocks, with reinforcement learning as the main approach and a tree search option too. It should be guided by a bioactivity prediction model I can train on my own data, then use that model to score building blocks and generate new molecules.

Please make it easy to install and run on a normal laptop, with optional GPU support if available. I also want a straightforward workflow for training the activity model, precomputing scores, generating molecules, and then filtering the results for novelty, predicted activity, and diversity. It should be flexible enough for me to swap in my own building blocks and reaction set later.

Include clear docs and runnable examples so I can reproduce the basic pipeline end to end. If anything is unclear, check the current docs online and make sensible defaults.