search

HaotianZhangAI4Science / Delete

Delete: Directly optimizing lead in protein pockets, including linker design, fragment elaboration, scaffold hopping and side-chain decoration
MIT License
22 stars 2 forks source link

Questions about how to train the model #7

Open zh2417 opened 6 months ago

zh2417 commented 6 months ago

Excellent work! I have some confusion. I would like to know if the final model obtained is a single "delete" model capable of executing all four subtasks of lead optimization (trained with seven masking strategies.?). Or does each subtask of lead optimization correspond to a separate "delete" model? (Each subtask model being trained with four masks, including three enhancement masks and one task-specific mask?) I noticed you mentioned only one checkpoint model. What subtask does this model correspond to?I look forward to your clarification. Thank you very much.

HaotianZhangAI4Science commented 3 months ago

Hi,

Yes, I only provided a ckpt for usage, but it is recommended to use task-specific ckpt to generate molecules, but you can also use mixed-training for obtaining a general ckpt.

I will open-source it as soon as possible! Thanks for your interest!

Best, Odin

zh2417 commented 3 months ago

So ，each subtask of lead optimization correspond to a separate "delete" model? (Each subtask model being trained with four masks, including three enhancement masks and one task-specific mask?) But the author, you only provided one model. Does this model correspond to the subtask of fragment growth?

zh2417 commented 3 months ago

Could you clarify if the model generates atoms one by one? Is it an autoregressive model? Does the generation of the next atom depend on the previously generated atoms?