Open chengthefang opened 6 years ago
@chf42 Hi, Thanks for your questions. In our paper, we try to design novel molecules with desired properties, which is formulated as black-box optimization problem. So, we combined MCTS with RNN for this problem. and using RNN to generate molecules, using MCTS for optimizing the properties. And you can understand it in this way, MCTS is used to guide the molecule generation of RNN. You can check the experiment result that we compared ChemTS with RNN-alone for optimizing logP, ChemTS can find higher scoring molecules. But RNN-alone is also good enough.
@yangxiufengsia Hi Xiufeng,
Thank you for your kind response. That makes more sense to me. I will try to play around with your model to see more functionality with that. Thank you again!
Cheng
Nice Paper!I have two questions:
Hi Xiufeng,
Nice paper published! I went through your paper and your codes, and found somewhere that me a little bit confused. i.e how to generate molecules in your model? via MCTS or RNN model, or both?
I wonder if it is possible to generate molecules using MCTS alone or using RNN model alone? If yes, how could I implement with your codes? Thank you!
Another piece I am confused it is that what's the exact difference between MCTS with RNN as rollout (this model) and the RNN model alone for molecular generation. In RNN-alone model, you can obtain the conditional probability for the next token, and sample on that or just pick up the one with maximum probability. That seems the same as how you select the next token in MCTS.
Thanks, Cheng