def _run_through(product_path, regressor, mol_chef_wae, seq_to_smi_list_func, cuda_details):
# == Get the products we want to retrosyntheize ==
with open(product_path, 'r') as fo:
products = [x.strip() for x in fo.readlines()]
# == Get graph representation of the ones we care about ==
graphs = []
for mol_smi in tqdm.tqdm(products, desc="creating graphs"):
mol = rdkit_general_ops.get_molecule(mol_smi, kekulize=True)
mol, am_to_indx_map = rdkit_general_ops.add_atom_mapping(mol)
mol_as_adj_list = rdkit_featurization_ops.mol_to_atom_feats_and_adjacency_list(mol, am_to_indx_map)
graph = atom_features_dataset.trfm_mol_as_adj_list_to_graph_as_adj_list_trsfm(mol_as_adj_list)
graphs.append(graph)
# == Now regress to latent space & run decoder ==
batch_size = 500
predicted_latents = []
resultant_reactants = []
for i in tqdm.tqdm(range(math.ceil(len(graphs) / batch_size)), desc="to_z_and_then_bag"):
graphs_of_batch = graphs[i*batch_size:(i+1)*batch_size]
graphs_of_batch = graphs_of_batch[0].concatenate(graphs_of_batch)
graphs_of_batch = graphs_of_batch.to_torch(cuda_details)
latents_ = regressor(graphs_of_batch)
predicted_latents.append(latents_.cpu().numpy())
seq_, _ = mol_chef_wae.decode_from_z_no_grad(latents_)
predicted_seqs_batch_first_np = seq_.cpu().numpy().T
for seq in predicted_seqs_batch_first_np:
seq_as_mols = seq_to_smi_list_func(seq)
reactant_str = '.'.join(sorted(seq_as_mols))
resultant_reactants.append(reactant_str)
return resultant_reactants
def transform_text_to_qed(text_line):
molecules = [rdkit_general_ops.get_molecule(mol_str, kekulize=False) for mol_str in text_line.split('.')]
qed_scores = [QED.qed(mol) for mol in molecules]
# May have many products so take max (given this is what we are optimising for in the optimisation part).
# Expect this to be less of an issue in practice as USPTO mostly details
# single product reactions. It may be interesting to look at using the Molecular Transformer prediction on
# these reactions rather than this ground truth and other ways of combining multiple products eg mean.
return np.max(qed_scores)
def __call__(self, path_to_smiles_file):
graphs = []
with open(path_to_smiles_file, 'r') as fo:
lines = fo.readlines()
for mol_smi in tqdm.tqdm(lines):
mol = rdkit_general_ops.get_molecule(mol_smi, kekulize=True)
mol, am_to_indx_map = rdkit_general_ops.add_atom_mapping(mol)
mol_as_adj_list = rdkit_featurization_ops.mol_to_atom_feats_and_adjacency_list(mol, am_to_indx_map)
graph = atom_features_dataset.trfm_mol_as_adj_list_to_graph_as_adj_list_trsfm(mol_as_adj_list)
graphs.append((graph,))
return graphs
def create_shared_dataset_files(reactants_to_reactant_id_dict):
print("creating shared files")
# Create file A
with open(
path.join(mchef_config.get_processed_data_dir(),
'reactants_to_reactant_id.json'), 'w') as fo:
json.dump(reactants_to_reactant_id_dict, fo)
# Create file B
print(f"Creating reactant smi to reactant_id map.")
reactant_feats = {}
for smiles, id in tqdm.tqdm(reactants_to_reactant_id_dict.items()):
mol = rdkit_general_ops.get_molecule(smiles, kekulize=True)
mol, am_to_indx_map = rdkit_general_ops.add_atom_mapping(mol)
reactant_feats[
id] = rdkit_featurization_ops.mol_to_atom_feats_and_adjacency_list(
mol, am_to_indx_map)
with open(
path.join(mchef_config.get_processed_data_dir(),
'reactants_feats.pick'), 'wb') as fo:
pickle.dump(reactant_feats, fo)
def _canonicalize(smi_str):
return rdkit_general_ops.return_canoncailised_smiles_str(
rdkit_general_ops.get_molecule(smi_str, kekulize=False),
kekuleSmiles=False)