Inquiry about FLOWSITE Model Inference Outputs

[L-Dandan]

Hi everyone,

I’ve been testing the FlowSite model and noticed that the ligand structures in the inference_out files sometimes appear chemically unreasonable—atoms are misconnected or placed in chaotic positions. Has anyone else encountered this issue?

Could the team or community kindly suggest:

Common pitfalls that might cause this (e.g., input formatting, missing parameters)?

Best practices to preprocess ligand inputs (e.g., file types, protonation states)?

Does this indicate a model limitation or a potential bug?

For reference, I’ve tried loading the optimal parameters provided or training the model myself, but they appear to be quite unusual.. Any advice would be greatly appreciated!

Thank you for your time and help.