Hi everyone,
I’m trying to better understand how target_hotspots are used during binder generation.
I found Issue #29 , where it was mentioned that the warning:
WARNING No hotspot_mask in batch, returning zeros for HotspotMaskSeqFeat
WARNING No hotspot_mask in batch, returning zeros for HotspotMaskPairFeat
comes from deprecated code and can generally be ignored because TargetConcatSeqFeat internally passes target_hotspot_mask as hotspot_mask.
However, I ran a simple experiment where I kept the target, search settings, and reward weights identical, and only changed the target_hotspots specification between two runs. The resulting outputs (reward CSVs and generated structures) were effectively identical. This made me wonder whether the hotspot specification is actually influencing generation in the current pipeline. While looking through the code, I noticed that:
- target_hotspot_mask is created and propagated through the dataset/search pipeline.
- The warning above is still emitted during generation.
- HotspotMaskSeqFeat and HotspotMaskPairFeat still look for batch["hotspot_mask"].
My question is:
Should changing target_hotspots be expected to change the generated binders in the current implementation? Or is hotspot conditioning currently inactive (or only affecting specific feature pathways) for the standard generation workflow?
I’m mainly trying to determine whether identical outputs after changing the hotspot set indicate a misunderstanding on my part, or whether this is expected behavior.
Thanks!
Hi everyone,
I’m trying to better understand how target_hotspots are used during binder generation.
I found Issue #29 , where it was mentioned that the warning:
WARNING No hotspot_mask in batch, returning zeros for HotspotMaskSeqFeat
WARNING No hotspot_mask in batch, returning zeros for HotspotMaskPairFeat
comes from deprecated code and can generally be ignored because TargetConcatSeqFeat internally passes target_hotspot_mask as hotspot_mask.
However, I ran a simple experiment where I kept the target, search settings, and reward weights identical, and only changed the target_hotspots specification between two runs. The resulting outputs (reward CSVs and generated structures) were effectively identical. This made me wonder whether the hotspot specification is actually influencing generation in the current pipeline. While looking through the code, I noticed that:
My question is:
Should changing target_hotspots be expected to change the generated binders in the current implementation? Or is hotspot conditioning currently inactive (or only affecting specific feature pathways) for the standard generation workflow?
I’m mainly trying to determine whether identical outputs after changing the hotspot set indicate a misunderstanding on my part, or whether this is expected behavior.
Thanks!