def __getitem__(self, index):
molecule_name = self.id[index]
graph_file = get_path(
) + 'data/graphs/' + self.graph_dir + '/%s.pickle' % molecule_name
graph = read_pickle_from_file(graph_file)
assert (graph.molecule_name == molecule_name)
mask = np.zeros(len(graph.coupling.type), np.bool)
for t in self.coupling_types:
mask += (graph.coupling.type == COUPLING_TYPE.index(t))
graph.coupling.id = graph.coupling.id[mask]
#graph.coupling.contribution = graph.coupling.contribution[mask]
graph.coupling.index = graph.coupling.index[mask]
graph.coupling.type = graph.coupling.type[mask]
graph.coupling.value = graph.coupling.value[mask]
atom = System(symbols=graph.axyz[0], positions=graph.axyz[1])
acsf = ACSF_GENERATOR.create(atom)
graph.node += [
acsf,
]
graph.node = np.concatenate(graph.node, -1)
graph.edge = np.concatenate(graph.edge, -1)
return graph
def __getitem__(self, index):
molecule_name = self.id[index]
graph_file = DATA_DIR + '/structure/graph1/%s.pickle' % molecule_name
graph = read_pickle_from_file(graph_file)
assert (graph.molecule_name == molecule_name)
# ##filter only J link
# if 0:
# # 1JHC, 2JHC, 3JHC, 1JHN, 2JHN, 3JHN, 2JHH, 3JHH
# mask = np.zeros(len(graph.coupling.type),np.bool)
# for t in ['1JHC', '2JHH']:
# mask += (graph.coupling.type == COUPLING_TYPE.index(t))
#
# graph.coupling.id = graph.coupling.id [mask]
# graph.coupling.contribution = graph.coupling.contribution [mask]
# graph.coupling.index = graph.coupling.index [mask]
# graph.coupling.type = graph.coupling.type [mask]
# graph.coupling.value = graph.coupling.value [mask]
if 1:
atom = System(symbols=graph.axyz[0], positions=graph.axyz[1])
acsf = ACSF_GENERATOR.create(atom)
graph.node += [
acsf,
]
# if 1:
# graph.edge = graph.edge[:-1]
graph.node = np.concatenate(graph.node, -1)
graph.edge = np.concatenate(graph.edge, -1)
return graph
def func_acsf(params):
i, molecule = params
#if i%1000 == 0:
# print(f"{i}th finish")
st = st_dict[molecule]
atoms = System(symbols=st["atom"].values,
positions=st[["x", "y", "z"]].values)
return gen.create(atoms)
def __getitem__(self, index):
molecule_name = self.id[index]
#graph_file = DATA_DIR + '/atoms-graph/graph/graph/%s.pickle'%molecule_name
#graph_file = DATA_DIR + '/graph-v4/graph_v4/graph_v4/%s.pickle'%molecule_name
graph_file = \
'../data/graph_v8/%s.pickle'%molecule_name
#graph_file = DATA_DIR + '/graph-v5/graph_v5/graph_v5/%s.pickle'%molecule_name
#graph_file = DATA_DIR + '/molecule-graph/graph_v2/graph_v2/%s.pickle'%molecule_name
graph = list(read_pickle_from_file(graph_file))
assert (graph[0] == molecule_name)
# ##filter only J link
# if 0:
# # 1JHC, 2JHC, 3JHC, 1JHN, 2JHN, 3JHN, 2JHH, 3JHH
# mask = np.zeros(len(graph.coupling.type),np.bool)
# for t in ['1JHC', '2JHH']:
# mask += (graph.coupling.type == COUPLING_TYPE.index(t))
#
# graph.coupling.id = graph.coupling.id [mask]
# graph.coupling.contribution = graph.coupling.contribution [mask]
# graph.coupling.index = graph.coupling.index [mask]
# graph.coupling.type = graph.coupling.type [mask]
# graph.coupling.value = graph.coupling.value [mask]
# add ACSF
atom = System(symbols=graph[2][0], positions=graph[2][1])
acsf = ACSF_GENERATOR.create(atom)
graph[3] += [
acsf,
]
graph[g_node_idx][7] = graph[g_node_idx][7].reshape([-1, 1])
graph[3] = np.concatenate(graph[3], -1)
dist = np.concatenate(graph[4], -1)[:, 4].reshape(-1, 1)
graph[4].append(1 / dist)
graph[4].append(1 / dist**2)
graph[4].append(1 / dist**3)
graph[4].append(1 / dist**6)
#for i in range(len(graph[4])):
# print(graph[4][])
graph[4] = np.concatenate(graph[4], -1)
graph[3][np.isnan(graph[3])] = 0
graph[4][np.isnan(graph[4])] = 0
# replace coupling atom_index2 -1 => 1
#if np.isnan(graph[3]).sum()>0 or np.isnan(graph[4]).sum() > 0:
# print(graph)
return graph
def get_scsf(data):
ret_list = []
for molecule_name in data["mol_names"]:
df = gb_structure.get_group(molecule_name)
df = df.sort_values(['atom_index'], ascending=True)
a = df.atom.values.tolist()
xyz = df[['x', 'y', 'z']].values
atom = System(symbols=a, positions=xyz)
acsf = ACSF_GENERATOR.create(atom)
acsf_df = pd.DataFrame(acsf)
acsf_df.columns = [f"acsf_{c}" for c in range(acsf_df.shape[1])]
acsf_df = pd.concat([
df[["molecule_name", "atom_index"]].reset_index(drop=True),
acsf_df.reset_index(drop=True)
],
axis=1)
ret_list.append(acsf_df)
return pd.concat(ret_list, axis=0)
def structure_to_graph(structure_file):
mol, smile = MolFromXYZ(structure_file)
factory = ChemicalFeatures.BuildFeatureFactory(os.path.join(RDConfig.RDDataDir, 'BaseFeatures.fdef'))
feature = factory.GetFeaturesForMol(mol)
structure = pd.read_csv(structure_file, skiprows=1, header=None, sep=" ",
names=["atom", "x", "y", "z"])
structure["radius"] = structure["atom"].map({'H': 0.38, 'C': 0.77, 'N': 0.75, 'O': 0.73, 'F': 0.71})
xyz = structure[["x", "y", "z"]]
norm_xyz = preprocessing.normalize(xyz, norm='l2')
n_atoms = mol.GetNumAtoms()
edge_array = []
bond_features = []
distance = []
rel_distance = []
angle = []
bond_vector = []
for i, j in itertools.product(range(n_atoms), repeat=2):
if i == j:
continue
edge_array.append((i, j))
bond = mol.GetBondBetweenAtoms(i, j)
if bond:
bond_type = bond.GetBondType()
else:
bond_type = None
bond_features.append(one_hot_encoding(bond_type, BONDS))
r = ((xyz.iloc[i] - xyz.iloc[j])**2).sum()**0.5
rel_dist = r/(structure.iloc[i]["radius"] +
structure.iloc[j]["radius"])
theta = (norm_xyz[i]*norm_xyz[j]).sum()
distance.append([r])
rel_distance.append([rel_dist]) # divide distance by sum of atomic radii
angle.append([theta])
bond_vector.append((xyz.iloc[i] - xyz.iloc[j]).tolist())
#distance = np.digitize(np.array(distance), bins=[0, 1, 2, 4, 8])
#rel_distance = np.digitize(np.array(rel_distance), bins=[0, 1, 2, 4, 8])
#angle = np.digitize(np.array(angle), bins=[-1, -.6, -.2, .2, .6])
edge_array = np.array(edge_array).T
edge_features = np.concatenate([
np.array(bond_features),
np.array(distance) / 4 - 1,
np.array(rel_distance) / 4 - 1,
np.array(angle), # absolute bond angle. Can use to calculate dihedral angles
np.array(bond_vector) # difference between coords of atoms i and j
], axis=1)
atom_features = defaultdict(list)
n_atoms = mol.GetNumAtoms()
for i in range(n_atoms):
atom = mol.GetAtomWithIdx(i)
atom_features["symbol"].append(one_hot_encoding(atom.GetSymbol(), SYMBOLS))
atom_features["aromatic"].append([atom.GetIsAromatic()])
atom_features["hybridization"].append(one_hot_encoding(atom.GetHybridization(), HYBRIDIZATIONS))
atom_features["num_h"].append([atom.GetTotalNumHs(includeNeighbors=True)])
atom_features["atomic"].append([atom.GetAtomicNum()])
atom = System(symbols=structure["atom"].values, positions=xyz.values)
acsf = ACSF_GENERATOR.create(atom)
atom_features["acsf"] = acsf
acceptor = np.zeros((n_atoms, 1), np.uint8)
donor = np.zeros((n_atoms, 1), np.uint8)
for feat in feature:
if feat.GetFamily() == 'Donor':
for i in feat.GetAtomIds():
donor[i] = 1
elif feat.GetFamily() == 'Acceptor':
for i in feat.GetAtomIds():
acceptor[i] = 1
print(len(atom_features["acsf"]), len(acceptor), len(donor))
atom_features = np.concatenate([atom_features["symbol"], atom_features["aromatic"],
atom_features["hybridization"], atom_features["num_h"],
atom_features["atomic"], atom_features["acsf"],
acceptor, donor], axis=1)
return edge_array, edge_features, atom_features, smile, xyz.values
def make_graph(name, gb_structure, gb_scalar_coupling):
# ['id', 'molecule_name', 'atom_index_0', 'atom_index_1', 'type','scalar_coupling_constant']
coupling_df = gb_scalar_coupling.get_group(name)
# [molecule_name,atom_index,atom,x,y,z]
df = gb_structure.get_group(name)
df = df.sort_values(['atom_index'], ascending=True)
a = df.atom.values.tolist()
xyz = df[['x', 'y', 'z']].values
mol = mol_from_axyz(a, xyz)
mol_op = openbabel.OBMol()
obConversion.ReadFile(mol_op, f'../input/champs-scalar-coupling/structures/{name}.xyz')
factory = ChemicalFeatures.BuildFeatureFactory(os.path.join(RDConfig.RDDataDir, 'BaseFeatures.fdef'))
feature = factory.GetFeaturesForMol(mol)
num_atom = mol.GetNumAtoms()
symbol = np.zeros((num_atom, len(SYMBOL)), np.uint8) # category
acceptor = np.zeros((num_atom, 1), np.uint8)
donor = np.zeros((num_atom, 1), np.uint8)
aromatic = np.zeros((num_atom, 1), np.uint8)
hybridization = np.zeros((num_atom, len(HYBRIDIZATION)), np.uint8)
num_h = np.zeros((num_atom, 1), np.float32) # real
atomic = np.zeros((num_atom, 1), np.float32)
# new features
degree = np.zeros((num_atom, 1), np.uint8)
formalCharge = np.zeros((num_atom, 1), np.float32)
chiral_tag = np.zeros((num_atom, 1), np.uint8)
crippen_contribs = np.zeros((num_atom, 2), np.float32)
tpsa = np.zeros((num_atom, 1), np.float32)
labute_asac = np.zeros((num_atom, 1), np.float32)
gasteiger_charges = np.zeros((num_atom, 1), np.float32)
esataindices = np.zeros((num_atom, 1), np.float32)
atomic_radiuss = np.zeros((num_atom, 1), np.float32)
electronegate = np.zeros((num_atom, 1), np.float32)
electronegate_sqre = np.zeros((num_atom, 1), np.float32)
mass = np.zeros((num_atom, 1), np.float32)
van = np.zeros((num_atom, 1), np.float32)
cov = np.zeros((num_atom, 1), np.float32)
ion = np.zeros((num_atom, 1), np.float32)
for i in range(num_atom):
atom = mol.GetAtomWithIdx(i)
atom_op = mol_op.GetAtomById(i)
symbol[i] = one_hot_encoding(atom.GetSymbol(), SYMBOL)
aromatic[i] = atom.GetIsAromatic()
hybridization[i] = one_hot_encoding(atom.GetHybridization(), HYBRIDIZATION)
num_h[i] = atom.GetTotalNumHs(includeNeighbors=True)
atomic[i] = atom.GetAtomicNum()
degree[i] = atom.GetTotalDegree()
formalCharge[i] = atom.GetFormalCharge()
chiral_tag[i] = int(atom.GetChiralTag())
crippen_contribs[i] = rdMolDescriptors._CalcCrippenContribs(mol)[i]
tpsa[i] = rdMolDescriptors._CalcTPSAContribs(mol)[i]
labute_asac[i] = rdMolDescriptors._CalcLabuteASAContribs(mol)[0][i]
gasteiger_charges[i] = atom_op.GetPartialCharge()
esataindices[i] = EState.EStateIndices(mol)[i]
atomic_radiuss[i] = atomic_radius[atom.GetSymbol()]
electronegate[i] = electronegativity[atom.GetSymbol()]
electronegate_sqre[i] = electronegativity_square[atom.GetSymbol()]
mass[i] = atomic_mass[atom.GetSymbol()]
van[i] = vanderwaalsradius[atom.GetSymbol()]
cov[i] = covalenzradius[atom.GetSymbol()]
ion[i] = ionization_energy[atom.GetSymbol()]
for t in range(0, len(feature)):
if feature[t].GetFamily() == 'Donor':
for i in feature[t].GetAtomIds():
donor[i] = 1
elif feature[t].GetFamily() == 'Acceptor':
for i in feature[t].GetAtomIds():
acceptor[i] = 1
num_edge = num_atom * num_atom - num_atom
edge_index = np.zeros((num_edge, 2), np.uint32)
bond_type = np.zeros((num_edge, len(BOND_TYPE)), np.uint32)
distance = np.zeros((num_edge, 1), np.float32)
angle = np.zeros((num_edge, 1), np.float32)
norm_xyz = preprocessing.normalize(xyz, norm='l2')
ij = 0
for i in range(num_atom):
for j in range(num_atom):
if i == j: continue
edge_index[ij] = [i, j]
bond = mol.GetBondBetweenAtoms(i, j)
if bond is not None:
bond_type[ij] = one_hot_encoding(bond.GetBondType(), BOND_TYPE)
distance[ij] = np.linalg.norm(xyz[i] - xyz[j])
angle[ij] = (norm_xyz[i] * norm_xyz[j]).sum()
ij += 1
xyz = xyz * 1.889726133921252
atom = System(symbols=a, positions=xyz)
acsf = ACSF_GENERATOR.create(atom)
l = []
for item in coupling_df[['atom_index_0', 'atom_index_1']].values.tolist():
i = edge_index.tolist().index(item)
l.append(i)
l = np.array(l)
coupling_edge_index = np.concatenate([coupling_df[['atom_index_0', 'atom_index_1']].values, l.reshape(len(l), 1)],
axis=1)
coupling = Coupling(coupling_df['id'].values,
coupling_df[['fc', 'sd', 'pso', 'dso']].values,
coupling_edge_index,
np.array([COUPLING_TYPE.index(t) for t in coupling_df.type.values], np.int32),
coupling_df['scalar_coupling_constant'].values,
)
graph = Graph(
name,
Chem.MolToSmiles(mol),
[a, xyz],
[acsf, symbol, acceptor, donor, aromatic, hybridization, num_h, atomic, degree, formalCharge, chiral_tag,
crippen_contribs, tpsa, labute_asac, gasteiger_charges, esataindices, atomic_radiuss, electronegate,
electronegate_sqre, mass, van, cov, ion],
[bond_type, distance, angle, ],
edge_index,
coupling,
)
return graph
def __getitem__(self, idx):
self.global_features = h5py.File(script_dir + '/../processed_data/global_116.h5', mode = 'r')
self.atom_features = h5py.File(script_dir + '/../processed_data/atom_116.h5', mode = 'r')
self.bond_features = h5py.File(script_dir + '/../processed_data/bond_116.h5', mode = 'r')
molecule_id = self.molecules_ids[idx]
molecule = self.molecules[idx]
atom_descriptor = self.atom_descriptors.loc[molecule_id]
bond_descriptor = self.bond_descriptors.loc[molecule_id]
#bond_descriptor = bond_descriptor.loc[bond_descriptor['bond_distance'] <= 3]
# Cycles
if molecule_id in self.cycles.index:
cycles = self.cycles.loc[molecule_id]
else:
cycles = pd.DataFrame(columns = self.cycles.columns)
cycles_edge_index = cycles['edge_index'].values.astype(np.int64)
cycles_id = cycles['cycle_id'].values.astype(np.int64)
# Edge connectivity
edges_connectivity = self.edges_connectivity.loc[molecule_id]
edges_connectivity_ids = np.copy(edges_connectivity[['edge_index_0', 'edge_index_1']].values.astype(np.int64).T)
edges_connectivity_vectors_0 = edges_connectivity[['vx_0', 'vy_0', 'vz_0']].values
edges_connectivity_vectors_1 = edges_connectivity[['vx_1', 'vy_1', 'vz_1']].values
edges_connectivity_feature_1 = np.sqrt(np.square(edges_connectivity_vectors_0).sum(axis = 1)).reshape(-1, 1)
edges_connectivity_feature_2 = np.sqrt(np.square(edges_connectivity_vectors_1).sum(axis = 1)).reshape(-1, 1)
edges_connectivity_feature_3 = np.sqrt(np.square(edges_connectivity_vectors_0 + edges_connectivity_vectors_1).sum(axis = 1)).reshape(-1, 1)
edges_connectivity_feature_0 = (edges_connectivity_vectors_0 * edges_connectivity_vectors_1).sum(axis = 1).reshape(-1, 1) / edges_connectivity_feature_1 / edges_connectivity_feature_2
edges_connectivity_features = np.concatenate([edges_connectivity_feature_0, edges_connectivity_feature_1, edges_connectivity_feature_2, edges_connectivity_feature_3], axis = 1)
atom = list(atom_descriptor['atom'])
xyz = atom_descriptor[['x', 'y', 'z']].values
xyz = apply_random_rotation(xyz)
connectivity = bond_descriptor[['atom_index_0', 'atom_index_1']].values
global_feature_numeric = np.copy(self.global_features['numeric'][molecule_id].reshape(1, -1))
global_feature_embeddings = np.copy(self.global_features['embeddings'][molecule_id].reshape(1, -1))
atom_indexes = atom_descriptor['index'].values
atom_index_min = atom_indexes.min()
atom_index_max = atom_indexes.max()
atom_feature_numeric = np.copy(self.atom_features['numeric'][atom_index_min : atom_index_max + 1][atom_indexes - atom_index_min])
atom_feature_embeddings = np.copy(self.atom_features['embeddings'][atom_index_min : atom_index_max + 1][atom_indexes - atom_index_min])
bond_indexes = bond_descriptor['index'].values
bond_index_min = bond_indexes.min()
bond_index_max = bond_indexes.max()
bond_feature_numeric = np.copy(self.bond_features['numeric'][bond_index_min : bond_index_max + 1][bond_indexes - bond_index_min])
bond_feature_embeddings = np.copy(self.bond_features['embeddings'][bond_index_min : bond_index_max + 1][bond_indexes - bond_index_min])
self.global_features.close()
self.atom_features.close()
self.bond_features.close()
# chemical descriptors
atom = System(symbols = atom, positions=xyz)
acsf = ACSF_GENERATOR.create(atom)
atom_feature_numeric = np.concatenate([atom_feature_numeric, xyz, acsf], axis = 1)
bond_vectors = build_bond_vector(connectivity, xyz)
bond_feature_numeric = np.concatenate([bond_feature_numeric, bond_vectors], axis = 1)
if self.name == 'train':
# Target
target = bond_descriptor['scalar_coupling_constant'].values.reshape(-1, 1)
target_mask = (bond_descriptor['type'] != 'VOID').values.reshape(-1, 1)
target_types = bond_descriptor['type_id'].values.reshape(-1, 1)
target_idx = bond_descriptor['edge_index'].values.reshape(-1, 1)
# data
data = Data(
x_numeric = torch.tensor(atom_feature_numeric, dtype = torch.float32),
x_embeddings = torch.tensor(atom_feature_embeddings, dtype = torch.int64),
edge_attr_numeric = torch.tensor(bond_feature_numeric, dtype = torch.float32),
edge_attr_embeddings = torch.tensor(bond_feature_embeddings, dtype = torch.int64),
u_numeric = torch.tensor(global_feature_numeric, dtype = torch.float32),
u_embeddings = torch.tensor(global_feature_embeddings, dtype = torch.int64),
edge_index = torch.tensor(connectivity.T),
num_nodes = atom_feature_numeric.shape[0],
molecule_ids = torch.tensor([molecule_id], dtype = torch.int64),
y = torch.tensor(target, dtype = torch.float32),
y_mask = torch.tensor(target_mask, dtype = torch.float32),
y_types = torch.tensor(target_types, dtype = torch.int64),
y_idx = torch.tensor(target_idx, dtype = torch.int32),
cycles_edge_index = torch.tensor(cycles_edge_index),
cycles_id = torch.tensor(cycles_id),
edges_connectivity_ids = torch.tensor(edges_connectivity_ids),
edges_connectivity_features = torch.tensor(edges_connectivity_features, dtype = torch.float32),
)
inputs = [
data.u_embeddings, data.x_embeddings, data.edge_attr_embeddings,
data.u_numeric, data.x_numeric, data.edge_attr_numeric,
]