Increasing num_diffn_steps to 500 increased total_inter_chain_clashes

[Ruibin-Liu]

I found that one chiral center of a ligand is inverted when I increased the num_diffn_timesteps from 100 to 500 and the total_inter_chain_clashes increased from 0 to 0.8. Is that expected? Is there a way to make sure the stereochemistry is always conserved in the difussion process?

[arogozhnikov]

single observation doesn't carry enough information to draw conclusions; you need to run at least several examples (and compare several seeds).

My prior: I don't expect more steps to provide worse results (on average, not in particular case), but you're welcome to test this experimentally

[Ruibin-Liu]

But I did observe that one chiral center is inverted, which is not desired for my case. I am not trying to assess the GENERAL'correctness' of Chai-1 but to know how to understand why that happened and how to avoid that.

[arogozhnikov]

Simply reject samples that don't satisfy your requirements (e.g. chirality).

why that happened and how to avoid that.

random sampling and model not following guidance. To get any more detailed/useful explanation you need to check multiple relevant examples.