Question on magnitude of pocket embedding change after HGNN refinement

[WenjinLiu2002]

Hi, thank you for releasing LigUnity — it’s a very impressive framework.

I’m currently exploring the protein pocket embedding refinement using the HGNN module. While testing the official demo data, I compared the pocket embedding before and after HGNN refinement using cosine similarity:
orig = orig.reshape(-1)
ref = refined.reshape(-1)
cos_sim = np.dot(orig, ref) / (np.linalg.norm(orig) * np.linalg.norm(ref))

For the demo example, I observed:
Cosine similarity ≈ 0.41
Cosine distance ≈ 0.59
This indicates a fairly large directional change in the pocket embedding after refinement.
I wanted to ask whether this magnitude of change is expected by design. Intuitively, I had assumed the HGNN refinement would slightly adjust the pocket embedding, whereas this result suggests a more substantial re-orientation in embedding space.

Could you help clarify:
Whether such a large cosine shift is typical for HGNN refinement?
How this level of change should be interpreted intuitively in the context of virtual screening?
How the refinement avoids over-biasing the pocket representation toward known ligand–pocket relationships?

Thank you again for your work.