can you get the coordinates directly?

Hi guys. According to your paper, maybe you guys had some difficulty getting the coordinates directly. I'm trying to modify your code for the docking pose prediction model to **make it a generative model** and also **directly output coordinates instead of the distance matrix**.

I'm thinking of mainly utilizing the **coordinate prediction part from the conformation model**. and also add the element prediction model from the conformation model as well. For the coordinate prediction part, i would also use the attention bias with the pocket atoms to update the masked ligand atoms. Have you guys tried this method?

there are some concerns here
1.the coordinate head seems equivariant, does that ensure the model is too? If it isn't equivariant, should we augment the data(reflection, translation etc.)
2.concerning the noise for the masked parts, i'm assuming that too much noise would make the model hard to predict. I wonder what you think of this part as well.
3. Maybe we can add an additional equivariant layer after getting the distances, that also takes the atom features
Thank you as always