difference of reporting experiment results

[yufengwhy]

There is difference between reporting experiment results of paper and results from github code.
Is the paper results from TBind v1.0.1, the github from TBind v0.5.0 ???

[luwei0917]

the result from the GitHub model should be about the same as the one reported in preprint. we only slightly updated the coordinates generation part of the code. The v1.0.1 is not released yet.

[yufengwhy]

@luwei0917
The model outputs the same affinity value for different structures of the same ligand: RDKit structure, PDBBind structure and predict structures of TankBind.
Is this normal?

[luwei0917]

Yes. In my opinion, that's the beauty of TankBind. You don't need prior knowledge of the complex structure to predict the affinity.

[yufengwhy]

@luwei0917
That sounds awesome.
I am curious about how tankbind achieve that the predicted affinity is conformation-agnostic, that is, the input of coordinates of ligands do not influence the outputted affinity;
If so, why can tankbind update these coordinates of ligands to improve the docked pose of ligands.

[luwei0917]

The coordinates of the ligand is determined based on its interaction with the protein and constraint imposed by chemical bonds and excluded volume effects.

[yufengwhy]

@luwei0917

1. predict coord
   D^c_{j,k} in Equation (5) is the group truth distance from PDBBind?
   If so, the label is leaked?
2. predict affinity
   D_{ij} in L_distance around Equation (4) is the group truth distance from PDBBind?
   If so, no matter what conformations input to the model, the best conformation can always be obtained, and the predict
   affinity is always based on the best conformation ?

[luwei0917]

It is based on the RDKit generated conformation, not PDBbind when testing or inferencing . It is based on PDBbind conformation during training.
A side note, even during training, under the self-dock setting, we also impose a LAS mask to the D^c_{j,k}.
2.
This is computing loss for training the model. We want our predicted D_{ij} to be as close to the ground truth as possible.
Affinity prediction is always based on the predicted interaction embedding.

The conformation of the ligand is predicted, not given.

[yufengwhy]

@luwei0917

1. "It is based on the RDKit generated conformation when testing or inferencing." This will make the conformation closer to RDKit, i.e., farther to PDB conformation？

[luwei0917]

The LAS mask exists in self dock. therefore, only the basic constraint is imposed. The general conformation is determined by the interaction.

[yufengwhy]

@luwei0917
Is TankBind SE(3)-equivariant？

[luwei0917]

yes. I think so. relative distance is SE(3)-equivariant.

[yufengwhy]

The LAS mask exists in self dock. therefore, only the basic constraint is imposed. The general conformation is determined by the interaction.

LAS includes rings, and the N-O-C ring in the predict example of your code, however, the N-O-C ring is not rigid, which should not in the LAS mask ?

[yufengwhy]

protein's GVP embedding and ligand's torchdrug feature, is these features invariant or equivariant?

[luwei0917]

I used "Chem.GetSymmSSSR" function in RDKit to get the ring. Could be that some ring should be rigid but considered flexible, or other way around. If you have better solution, please let me know.

[yufengwhy]

I used "Chem.GetSymmSSSR" function in RDKit to get the ring. Could be that some ring should be rigid but considered flexible, or other way around. If you have better solution, please let me know.

Chem.GetSymmSSSR is used to obtain symmetry smallest ring, not necessarily rigid. So we could choose from "Chem.GetSymmSSSR" rings that have only one RDKit conformation?