Encode molecules starting from their SMILES

[marcosbodio]

Hello, I would like to know if it is possible to use GraphMVP to encode molecule starting from their SMILES. I have read this issue, but that does not help much. I would be really grateful if you could provide some explanation, and ideally an example. Thank you!

[chao1224]

Hi @marcosbodio,

Thank you for your question.

• SMILES is a string representation for the molecule topology, and is not exactly the same as the 2D graph.
• So if we want to reuse the implementation of the current repo, then the answer is no.
• However, if we expand GraphMVP from 2D-3D to topology-geometry, then the answer is yes. What you need to do is to replace the 2D graph + 2D GNN (GIN in our paper) with SMILES + BERT (or any other sequence encoder).

[marcosbodio]

Hi @chao1224, thank you for your answer. I see in your paper that you have Table 5 where you list results on DTA tasks with Davis and KIBA. These datasets contains SMILES of molecules, so how did you use GraphMVP (or GraphMVP-G, GraphMVP-C) on these datasets? It would be very useful to see the code, because that would clarify what is the proper way of using your model starting from the SMILES of molecule.

[chao1224]

Hi @marcosbodio,

Sure, you can check this python script, specifically, this line assigns which dataset to use.

[marcosbodio]

Hi @chao1224, I have looked at the script that you linked above, and I think that is for fine tuning your model, which I would prefer to avoid.

I was hoping to use a checkpoint of your model, for example output/3D_hybrid_02_masking/GEOM_3D_nmol50000_nconf5_nupper1000/CL_1_VAE_1/6_51_10_0.1/0.3_EBM_dot_prod_0.1_normalize_l2_detach_target_2_100_0/pretraining_model.pth in GraphMVP_simple_features_for_classification.zip (shared here)

I wonder if I could do something like this:

import torch
from rdkit import Chem
from rdkit.Chem.rdDistGeom import EmbedMolecule

from src_classification.GEOM_dataset_preparation import mol_to_graph_data_obj_simple_3D

smiles = 'Cn1cnc(c1)C(=O)c1ccc(CN2[C@H](Cc3ccccn3)C(=O)Nc3cc(Cl)ccc3C2=O)cc1'
mol = Chem.MolFromSmiles(smiles)
mol = Chem.AddHs(mol)
EmbedMolecule(mol=mol)
data = mol_to_graph_data_obj_simple_3D(mol)

and then feed data to the model loaded from the checkpoint to compute an embedding of the SMILES. What do you think?

[chao1224]

Hi @marcosbodio,

Yes, I think this is right if you want to use the 3D representation.

1. When we create the checkpoints, we save the following modules (code):

            saver_dict = {
                'model': molecule_model_2D.state_dict(),
                'model_3D': molecule_model_3D.state_dict(),
                'AE_2D_3D_model': AE_2D_3D_model.state_dict(),
                'AE_3D_2D_model': AE_3D_2D_model.state_dict(),
            }

2. What you wrote above can be fed into the model_3D.
3. If you only want to use the 2D checkpoint, which is model above, then you can follow this pseudocode:

smiles = 'Cn1cnc(c1)C(=O)c1ccc(CN2[C@H](Cc3ccccn3)C(=O)Nc3cc(Cl)ccc3C2=O)cc1'
mol = Chem.MolFromSmiles(smiles)

data = mol_to_graph_data_obj_simple(mol)

where mol_to_graph_data_obj_simple is in this function.

[marcosbodio]

HI @chao1224 ,

I have tried to load one of your model checkpoint, but I do not see model_3D. Here is what I did:

model_path = 'output/3D_hybrid_02_masking/GEOM_3D_nmol50000_nconf5_nupper1000/CL_1_VAE_1/6_51_10_0.1/0.3_EBM_dot_prod_0.1_normalize_l2_detach_target_2_100_0/pretraining_model.pth'
model = torch.load(f=model_path, map_location=torch.device('cpu'))
print(model.keys())
print('model_3D' in model)

where model_path is from your file GraphMVP_simple_features_for_classification.zip (shared here)

The previous code prints the following:

odict_keys(['x_embedding1.weight', 'x_embedding2.weight', 'gnns.0.mlp.0.weight', 'gnns.0.mlp.0.bias', 'gnns.0.mlp.2.weight', 'gnns.0.mlp.2.bias', 'gnns.0.edge_embedding1.weight', 'gnns.0.edge_embedding2.weight', 'gnns.1.mlp.0.weight', 'gnns.1.mlp.0.bias', 'gnns.1.mlp.2.weight', 'gnns.1.mlp.2.bias', 'gnns.1.edge_embedding1.weight', 'gnns.1.edge_embedding2.weight', 'gnns.2.mlp.0.weight', 'gnns.2.mlp.0.bias', 'gnns.2.mlp.2.weight', 'gnns.2.mlp.2.bias', 'gnns.2.edge_embedding1.weight', 'gnns.2.edge_embedding2.weight', 'gnns.3.mlp.0.weight', 'gnns.3.mlp.0.bias', 'gnns.3.mlp.2.weight', 'gnns.3.mlp.2.bias', 'gnns.3.edge_embedding1.weight', 'gnns.3.edge_embedding2.weight', 'gnns.4.mlp.0.weight', 'gnns.4.mlp.0.bias', 'gnns.4.mlp.2.weight', 'gnns.4.mlp.2.bias', 'gnns.4.edge_embedding1.weight', 'gnns.4.edge_embedding2.weight', 'batch_norms.0.weight', 'batch_norms.0.bias', 'batch_norms.0.running_mean', 'batch_norms.0.running_var', 'batch_norms.0.num_batches_tracked', 'batch_norms.1.weight', 'batch_norms.1.bias', 'batch_norms.1.running_mean', 'batch_norms.1.running_var', 'batch_norms.1.num_batches_tracked', 'batch_norms.2.weight', 'batch_norms.2.bias', 'batch_norms.2.running_mean', 'batch_norms.2.running_var', 'batch_norms.2.num_batches_tracked', 'batch_norms.3.weight', 'batch_norms.3.bias', 'batch_norms.3.running_mean', 'batch_norms.3.running_var', 'batch_norms.3.num_batches_tracked', 'batch_norms.4.weight', 'batch_norms.4.bias', 'batch_norms.4.running_mean', 'batch_norms.4.running_var', 'batch_norms.4.num_batches_tracked'])
False

Am I loading the wrong checkpoint?

[chao1224]

Hi @marcosbodio ,

I need to double-check the checkpoint files when I got time. Meanwhile, you should be able to use this checkpoint, which is one of the SOTA PaiNN pretraining methods (paper link)).