Questions about benchmarking Hit2Lead algorithm

[Kartinaa]

Dear authors,

Hope this message finds you well. Thanks so much for sharing such a valuable resource with people in the area to standardize the benchmarking process of SBGMs.

I have no issues with the de novo algorithms benchmarking. However, I have some questions regarding the h2l models, and I’d like to make sure my understanding is correct. Specifically, for diffusion-based models such as Delete and DiffDec, my understanding is that they require a scaffold as input, which the model then grows or decorates.

In your README, however, you mention that the input is a full molecule rather than a scaffold. In that case, I’m wondering how the generation process is handled, as it seems challenging for diffusion models to regenerate known actives when conditioned on an entire molecule. Could you please clarify how generation is performed for these models in your benchmark setting?

Many thanks for your time and help. I really appreciate the effort you put into making this resource available, and I hope you have a great day!

[HShokaku]

Hi @Kartinaa ,

Sorry for the confusion caused by the earlier version of the README. We have recently updated the README to clarify the H2L setup. To make this concrete and reproducible, we now provide an explicit example script get_h2l_scaf.py.

The extracted scaffold molecules are used as the inputs to these H2L models for subsequent growth or decoration.

[Kartinaa]

Hello @HShokaku ,

Thanks so much for your kind reply! I'll take a look see what happen!