Regarding the reference coordinate issue in the Dim6RotStructureHead of the structure decoder？

[lonelycrab888]

Hello,

I noticed that in the paper, during Stage 2 training of the structure tokenizer, full-atom modeling is performed. However, it seems that the provided repository code does not include this functionality.

I would like to ask whether, in Dim6RotStructureHead, achieving full-atom (atom14) output requires additional processing of angles and modifications to the reference coordinates (BB_COORDINATES). Could you kindly provide some guidance on how to adjust the reference coordinates?

Additionally, could you explain how rot_frames(sin_theta, cos_theta) in Algorithm 8 of the paper is implemented?

Thank you!