Partial diffusion / optimisation without explicit objective

[adrienchaton]

Hello and thanks for sharing your model,

I would like to evaluate its ability to "reconstruct" known binders for a given target pocket. (or at least recreate similar interactions)
For that I would want to apply some noising steps to the reference ligands and then ask the model to denoise these steps and compare what molecules we get compared to the known binders.

I guessed it is similar to the optimisation workflow you implement, but without explicit objective to optimise for which may bias the model to other chemical spaces than the reference ligands. When I check https://github.com/arneschneuing/DiffSBDD/blob/main/optimize.py it seems that the objective is only a post-scoring and that the model sampling is not biased towards specific properties... is that right?

So if I want to do my experiment, may I just run optimize.py regardless of the objective and ignore the corresponding score?
https://github.com/arneschneuing/DiffSBDD/blob/main/optimize.py#L239

From the configs, I see that model were trained on 500 diffusion steps, which I assume is the default for de-novo generation.
Would you also use 100 noising/denoising steps to evaluate your model reconstruction abilities as is the default in the opitimize workflow?

Thanks!