Example to add ESOL dataset

[thegodone]

You give an interesting code and example, but I don't see how to use it with esol dataset.

I have seen that you combine pos and atom label in your data.x. but how do you build your edge_index ?

is it a local bonded index like in classical graphs or for full molecule indexes ?

Also does the Hs are needed for other target than QM9 ?

[zetayue]

I have seen that you combine pos and atom label in your data.x. but how do you build your edge_index ?
is it a local bonded index like in classical graphs or for full molecule indexes ?

We represent each molecular graph as a multiplex graph including (1) the global layer that captures global interactions and (2) the local layer that captures local interactions. Thus our model needs two edge_index information, edge_index_g and edge_index_l:

• For the global layer, edge_index_g is created by setting a cutoff distance via the radius function provided by torch_geometric:

  MXMNet/model.py

  Lines 108 to 109 in 1d31517

  	row, col = radius(pos, pos, self.cutoff, batch, batch, max_num_neighbors=500)
  	edge_index_g = torch.stack([row, col], dim=0)

• For the local layer, edge_index_l can be achieved by using either the chemical bond information or using a cutoff distance. In our code, we use the chemical bond information to build edge_index_l via the way provided by torch_geometric: https://github.com/rusty1s/pytorch_geometric/blob/8b89705072e11da058e98f180b062b28d0939de5/torch_geometric/datasets/qm9.py#L233-L246

Also does the Hs are needed for other target than QM9 ?

We think this depends on the task being investigated as well as the computational resources. Besides, if the 3D molecular structures are not precise enough, including Hs might not be a good idea.