def get_USRlike_atoms(self):
"""Returns 4 rdkit Point3D objects similar to those used in USR:
- centroid (ctd)
- closest to ctd (cst)
- farthest from cst (fct) (usually ctd but let's avoid computing too many dist matrices)
- farthest from fct (ftf)"""
matrix = rdmolops.Get3DDistanceMatrix(self.mol)
conf = self.mol.GetConformer()
coords = conf.GetPositions()
# centroid
ctd = rdMolTransforms.ComputeCentroid(conf)
# closest to centroid
min_dist = 100
for atom in self.mol.GetAtoms():
point = rdGeometry.Point3D(*coords[atom.GetIdx()])
dist = ctd.Distance(point)
if dist < min_dist:
min_dist = dist
cst = point
cst_idx = atom.GetIdx()
# farthest from cst
fct_idx = argmax(matrix[cst_idx])
fct = rdGeometry.Point3D(*coords[fct_idx])
# farthest from fct
ftf_idx = argmax(matrix[fct_idx])
ftf = rdGeometry.Point3D(*coords[ftf_idx])
return ctd, cst, fct, ftf
def construct_distance_matrix(mol, out_size=-1, contain_Hs=False):
"""Construct distance matrix
Args:
mol (Chem.Mol):
out_size (int):
contain_Hs (bool):
Returns (numpy.ndarray): 2 dimensional array which represents distance
between atoms
"""
if mol is None:
raise MolFeatureExtractionError('mol is None')
N = mol.GetNumAtoms()
if out_size < 0:
size = N
elif out_size >= N:
size = out_size
else:
raise MolFeatureExtractionError('out_size {} is smaller than number '
'of atoms in mol {}'.format(
out_size, N))
if contain_Hs:
mol2 = mol
else:
mol2 = AllChem.AddHs(mol)
conf_id = AllChem.EmbedMolecule(mol2)
if not contain_Hs:
mol2 = AllChem.RemoveHs(mol2)
try:
dist_matrix = rdmolops.Get3DDistanceMatrix(mol2, confId=conf_id)
except ValueError as e:
logger = getLogger(__name__)
logger.info('construct_distance_matrix failed, type: {}, {}'.format(
type(e).__name__, e.args))
logger.debug(traceback.format_exc())
raise MolFeatureExtractionError
if size > N:
dists = numpy.zeros((size, size), dtype=numpy.float32)
a0, a1 = dist_matrix.shape
dists[:a0, :a1] = dist_matrix
else:
dists = dist_matrix
return dists.astype(numpy.float32)
def construct_distance_matrix(mol, out_size=-1):
"""Construct distance matrix
Args:
mol (Chem.Mol):
out_size (int):
Returns:
"""
if mol is None:
raise MolFeatureExtractionError('mol is None')
N = mol.GetNumAtoms()
if out_size < 0:
size = N
elif out_size >= N:
size = out_size
else:
raise MolFeatureExtractionError('out_size {} is smaller than number '
'of atoms in mol {}'.format(
out_size, N))
confid = AllChem.EmbedMolecule(mol)
try:
dist_matrix = rdmolops.Get3DDistanceMatrix(mol, confId=confid)
except ValueError as e:
logger = getLogger(__name__)
logger.info('construct_distance_matrix failed, type: {}, {}'.format(
type(e).__name__, e.args))
logger.debug(traceback.format_exc())
raise MolFeatureExtractionError
if size > N:
dists = numpy.zeros((size, size), dtype=numpy.float32)
a0, a1 = dist_matrix.shape
dists[:a0, :a1] = dist_matrix
else:
dists = dist_matrix
return dists.astype(numpy.float32)
def construct_pair_feature(mol, use_all_feature):
"""construct pair feature
Args:
mol (Mol): mol instance
use_all_feature (bool):
If True, all pair features are extracted.
If False, a part of pair features is extracted.
You can confirm the detail in the paper.
Returns:
features (numpy.ndarray): The shape is (num_edges, num_edge_features)
bond_idx (numpy.ndarray): The shape is (2, num_edges)
bond_idx[0] represents the list of StartNodeIdx and bond_idx[1]
represents the list of EndNodeIdx.
"""
converter = GaussianDistance()
# prepare the data for extracting the pair feature
bonds = mol.GetBonds()
graph_distance_matrix = Chem.GetDistanceMatrix(mol)
is_in_ring = get_is_in_ring(mol)
confid = AllChem.EmbedMolecule(mol)
try:
coordinate_matrix = rdmolops.Get3DDistanceMatrix(
mol, confId=confid)
except ValueError as e:
logger = getLogger(__name__)
logger.info('construct_distance_matrix failed, type: {}, {}'
.format(type(e).__name__, e.args))
logger.debug(traceback.format_exc())
raise MolFeatureExtractionError
feature = []
bond_idx = []
for bond in bonds:
start_node = bond.GetBeginAtomIdx()
end_node = bond.GetEndAtomIdx()
# create pair feature
distance_feature = numpy.array(
graph_distance_matrix[start_node][end_node], dtype=numpy.float32)
bond_feature = construct_bond_vec(mol, start_node, end_node)
ring_feature = construct_ring_feature_vec(
is_in_ring, start_node, end_node)
bond_idx.append((start_node, end_node))
if use_all_feature:
expanded_distance_feature = \
construct_expanded_distance_vec(
coordinate_matrix, converter, start_node, end_node)
feature.append(numpy.hstack((bond_feature, ring_feature,
distance_feature,
expanded_distance_feature)))
else:
expanded_distance_feature = \
construct_expanded_distance_vec(
coordinate_matrix, converter, start_node, end_node)
feature.append(expanded_distance_feature)
bond_idx = numpy.array(bond_idx).T
feature = numpy.array(feature)
return feature, bond_idx
def perform_lrp(model, hyper, trial=0, sample=None, epsilon=0.1, gamma=0.1):
tf.config.experimental.set_memory_growth(
tf.config.experimental.list_physical_devices('GPU')[0], True)
# Make folder
fig_path = "../analysis/{}".format(model)
if not os.path.isdir(fig_path):
os.mkdir(fig_path)
fig_path = "../analysis/{}/{}".format(model, hyper)
if not os.path.isdir(fig_path):
os.mkdir(fig_path)
fig_path = "../analysis/{}/{}/heatmap".format(model, hyper)
if not os.path.isdir(fig_path):
os.mkdir(fig_path)
# Load results
base_path = "../result/{}/{}/".format(model, hyper)
path = base_path + 'trial_{:02d}/'.format(trial)
# Load hyper
with open(path + 'hyper.csv', newline='') as csvfile:
reader = csv.DictReader(csvfile)
for row in reader:
hyper = dict(row)
# Load model
custom_objects = {
'NodeEmbedding': NodeEmbedding,
'GraphConvolution': GraphConvolution,
'Normalize': Normalize,
'GlobalPooling': GlobalPooling
}
model = load_model(path + 'best_model.h5', custom_objects=custom_objects)
print([l.name for l in model.layers])
# Load data
data = np.load(path + 'data_split.npz')
dataset = Dataset('refined', 5)
if sample is not None:
dataset.split_by_idx(32, data['train'], data['valid'],
data['test'][sample])
else:
dataset.split_by_idx(32, data['train'], data['valid'], data['test'])
data.close()
# Predict
true_y = dataset.test_y
outputs = {}
for layer_name in [
'node_embedding', 'node_embedding_1', 'normalize', 'normalize_1',
'activation', 'add', 'activation_1', 'add_1', 'global_pooling',
'activation_2', 'activation_3', 'activation_4',
'atom_feature_input'
]:
sub_model = tf.keras.models.Model(
inputs=model.input, outputs=model.get_layer(layer_name).output)
outputs[layer_name] = sub_model.predict(dataset.test,
steps=dataset.test_step,
verbose=0)[:len(true_y)]
# Output layer: LRP-0
# print('Calculating Dense_2...')
relevance = lrp_dense(outputs['activation_3'],
outputs['activation_4'],
model.get_layer('dense_2').get_weights()[0],
model.get_layer('dense_2').get_weights()[1],
epsilon=0)
# Dense layer: LRP-e
# print('Calculating Dense_1...')
relevance = lrp_dense(outputs['activation_2'],
relevance,
model.get_layer('dense_1').get_weights()[0],
model.get_layer('dense_1').get_weights()[1],
epsilon=epsilon)
# Dense layer: LRP-e
# print('Calculating Dense_0...')
relevance = lrp_dense(outputs['global_pooling'],
relevance,
model.get_layer('dense').get_weights()[0],
model.get_layer('dense').get_weights()[1],
epsilon=epsilon)
# Pooling layer
# print('Calculating Pooling...')
relevance = lrp_pooling(outputs['activation_1'], relevance)
# Add layer
# print('Calculating Add_1...')
relevance_1, relevance_2 = lrp_add(
[outputs['add'], outputs['activation_1']], relevance)
# GCN layer: LRP-g
# print('Calculating GCN_1...')
relevance = lrp_gcn_gamma(
outputs['add'],
relevance_2,
outputs['normalize_1'],
model.get_layer('graph_convolution_1').get_weights()[0],
gamma=gamma) + relevance_1
# Add layer
# print('Calculating Add_0...')
relevance_1, relevance_2 = lrp_add(
[outputs['graph_embedding_1'], outputs['activation']], relevance)
# GCN layer: LRP-g
# print('Calculating GCN_0...')
relevance = lrp_gcn_gamma(
outputs['graph_embedding_1'],
relevance_2,
outputs['normalize'],
model.get_layer('graph_convolution').get_weights()[0],
gamma=gamma) + relevance_1
# Embedding layer : LRP-e
# print('Calculating Embedding_1...')
relevance = lrp_dense(
outputs['graph_embedding'],
relevance,
model.get_layer('graph_embedding_1').get_weights()[0],
model.get_layer('graph_embedding_1').get_weights()[1],
epsilon=epsilon)
# Embedding layer : LRP-e
# print('Calculating Embedding_0...')
relevance = lrp_dense(outputs['atom_feature_input'],
relevance,
model.get_layer('graph_embedding').get_weights()[0],
model.get_layer('graph_embedding').get_weights()[1],
epsilon=epsilon)
relevance = tf.math.reduce_sum(relevance, axis=-1).numpy()
relevance = np.divide(relevance, np.expand_dims(true_y, -1))
# Preset
DrawingOptions.bondLineWidth = 1.5
DrawingOptions.elemDict = {}
DrawingOptions.dotsPerAngstrom = 20
DrawingOptions.atomLabelFontSize = 4
DrawingOptions.atomLabelMinFontSize = 4
DrawingOptions.dblBondOffset = 0.3
DrawingOptions.wedgeDashedBonds = False
# Load data
dataframe = pd.read_pickle('../data/5A.pkl')
if sample is not None:
test_set = np.load(path + 'data_split.npz')['test'][sample]
else:
test_set = np.load(path + 'data_split.npz')['test']
# Draw images for test molecules
colormap = cm.get_cmap('seismic')
for idx, test_idx in enumerate(test_set):
print('Drawing figure for {}/{}'.format(idx, len(test_set)))
pdb_code = dataframe.iloc[test_idx]['code']
error = np.absolute(dataframe.iloc[test_idx]['output'] -
outputs['activation_4'][idx])[0]
if error > 0.2: continue
for mol_ligand, mol_pocket in zip(
Chem.SDMolSupplier(
'../data/refined-set/{}/{}_ligand.sdf'.format(
pdb_code, pdb_code)),
Chem.SDMolSupplier(
'../data/refined-set/{}/{}_pocket.sdf'.format(
pdb_code, pdb_code))):
# Crop atoms
mol = Chem.CombineMols(mol_ligand, mol_pocket)
distance = np.array(rdmolops.Get3DDistanceMatrix(mol))
cropped_idx = np.argwhere(
np.min(distance[:, :mol_ligand.GetNumAtoms()], axis=1) <= 5
).flatten()
unpadded_relevance = np.zeros((mol.GetNumAtoms(), ))
np.put(unpadded_relevance, cropped_idx, relevance[idx])
scale = max(max(unpadded_relevance),
math.fabs(min(unpadded_relevance))) * 3
# Separate fragments in Combined Mol
idxs_frag = rdmolops.GetMolFrags(mol)
mols_frag = rdmolops.GetMolFrags(mol, asMols=True)
# Draw fragment and interaction
for i, (mol_frag,
idx_frag) in enumerate(zip(mols_frag[1:], idxs_frag[1:])):
# Ignore water
if mol_frag.GetNumAtoms() == 1:
continue
# Generate 2D image
mol_combined = Chem.CombineMols(mols_frag[0], mol_frag)
AllChem.Compute2DCoords(mol_combined)
fig = Draw.MolToMPL(mol_combined, coordScale=1)
fig.axes[0].set_axis_off()
# Draw line between close atoms (5A)
flag = False
for j in range(mol_ligand.GetNumAtoms()):
for k in idx_frag:
if distance[j, k] <= 5:
# Draw connection
coord_li = mol_combined._atomPs[j]
coord_po = mol_combined._atomPs[
idx_frag.index(k) + mols_frag[0].GetNumAtoms()]
x, y = np.array([[coord_li[0], coord_po[0]],
[coord_li[1], coord_po[1]]])
line = Line2D(x,
y,
color='b',
linewidth=1,
alpha=0.3)
fig.axes[0].add_line(line)
flag = True
# Draw heatmap for atoms
for j in range(mol_combined.GetNumAtoms()):
relevance_li = unpadded_relevance[j]
relevance_li = relevance_li / scale + 0.5
highlight = plt.Circle(
(mol_combined._atomPs[j][0],
mol_combined._atomPs[j][1]),
0.035 * math.fabs(unpadded_relevance[j] / scale) +
0.008,
color=colormap(relevance_li),
alpha=0.8,
zorder=0)
fig.axes[0].add_artist(highlight)
# Save
if flag:
fig_name = fig_path + '/{}_lrp_{}_{}_{}.png'.format(
trial, test_idx, pdb_code, i)
fig.savefig(fig_name, bbox_inches='tight')
plt.close(fig)
def parse_dataset(self):
def _one_hot(x, allowable_set):
return list(map(lambda s: x == s, allowable_set))
# Get total types of atoms
for ligand, pocket in zip(self.x_ligand, self.x_pocket):
mol = Chem.CombineMols(ligand, pocket)
self.num_atoms = max(self.num_atoms, mol.GetNumAtoms())
for atom in mol.GetAtoms():
symbol = atom.GetSymbol()
if symbol not in self.atom_type.keys():
self.atom_type[symbol] = 1
else:
self.atom_type[symbol] += 1
self.atom_type = {
k: v
for k, v in sorted(
self.atom_type.items(), key=lambda item: item[1], reverse=True)
}
columns = [
'code', 'symbol', 'atomic_num', 'degree', 'hybridization',
'implicit_valence', 'formal_charge', 'aromaticity', 'ring_size',
'num_hs', 'acid_base', 'h_donor_acceptor', 'adjacency_intra',
'adjacency_inter', 'distance', 'output'
]
self.dataframe = pd.DataFrame(columns=columns)
hydrogen_donor = Chem.MolFromSmarts(
"[$([N;!H0;v3,v4&+1]),$([O,S;H1;+0]),n&H1&+0]")
hydrogen_acceptor = Chem.MolFromSmarts(
"[$([O,S;H1;v2;!$(*-*=[O,N,P,S])]),$([O,S;H0;v2]),$([O,S;-]),$([N;v3;!$(N-*=[O,N,P,S])]),n&H0&+0,$([o,s;+0;!$([o,s]:n);!$([o,s]:c:n)])]"
)
acidic = Chem.MolFromSmarts("[$([C,S](=[O,S,P])-[O;H1,-1])]")
basic = Chem.MolFromSmarts(
"[#7;+,$([N;H2&+0][$([C,a]);!$([C,a](=O))]),$([N;H1&+0]([$([C,a]);!$([C,a](=O))])[$([C,a]);!$([C,a](=O))]),$([N;H0&+0]([C;!$(C(=O))])([C;!$(C(=O))])[C;!$(C(=O))])]"
)
for ligand, pocket in zip(self.x_ligand, self.x_pocket):
n_ligand = ligand.GetNumAtoms()
mol = Chem.CombineMols(ligand, pocket)
# Crop atoms
adjacency = np.array(rdmolops.Get3DDistanceMatrix(mol))
idx = np.argwhere(
np.min(adjacency[:, :n_ligand], axis=1) <= self.cutoff
).flatten().tolist()
# Get tensors
Chem.AssignStereochemistry(mol)
hydrogen_donor_match = sum(mol.GetSubstructMatches(hydrogen_donor),
())
hydrogen_acceptor_match = sum(
mol.GetSubstructMatches(hydrogen_acceptor), ())
acidic_match = sum(mol.GetSubstructMatches(acidic), ())
basic_match = sum(mol.GetSubstructMatches(basic), ())
ring = mol.GetRingInfo()
m = [[], [], [], [], [], [], [], [], [], [], [], [], [], [], [],
[]]
m[0] = ligand.GetProp('_Name').split('_')[0]
for atom_idx in idx:
atom = mol.GetAtomWithIdx(atom_idx)
m[1].append(_one_hot(atom.GetSymbol(), self.atom_type.keys()))
m[2].append([atom.GetAtomicNum()])
m[3].append(_one_hot(atom.GetDegree(), [0, 1, 2, 3, 4, 5, 6]))
m[4].append(
_one_hot(atom.GetHybridization(), [
Chem.rdchem.HybridizationType.SP,
Chem.rdchem.HybridizationType.SP2,
Chem.rdchem.HybridizationType.SP3,
Chem.rdchem.HybridizationType.SP3D,
Chem.rdchem.HybridizationType.SP3D2
]))
m[5].append(
_one_hot(atom.GetImplicitValence(), [0, 1, 2, 3, 4, 5, 6]))
m[6].append(
_one_hot(atom.GetFormalCharge(), [-3, -2, -1, 0, 1, 2, 3]))
m[7].append([atom.GetIsAromatic()])
m[8].append([
ring.IsAtomInRingOfSize(atom_idx, 3),
ring.IsAtomInRingOfSize(atom_idx, 4),
ring.IsAtomInRingOfSize(atom_idx, 5),
ring.IsAtomInRingOfSize(atom_idx, 6),
ring.IsAtomInRingOfSize(atom_idx, 7),
ring.IsAtomInRingOfSize(atom_idx, 8)
])
m[9].append(_one_hot(atom.GetTotalNumHs(), [0, 1, 2, 3, 4]))
m[10].append(
[atom_idx in acidic_match, atom_idx in basic_match])
m[11].append([
atom_idx in hydrogen_donor_match, atom_idx
in hydrogen_acceptor_match
])
m[12] = np.array(rdmolops.GetAdjacencyMatrix(mol))[idx][:, idx]
adj = np.zeros_like(m[12])
adj[:n_ligand, n_ligand:] = 1.
adj[n_ligand:, :n_ligand] = 1.
m[13] = adj
m[14] = np.array(rdmolops.Get3DDistanceMatrix(mol))[idx][:, idx]
m[15] = float(ligand.GetProp('target_calc'))
self.dataframe = self.dataframe.append(pd.DataFrame(
[m], columns=columns),
ignore_index=True,
sort=True)
# Pad data
self.num_atoms = 0
for i in range(len(self.dataframe)):
self.num_atoms = max(len(self.dataframe.iloc[i]['symbol']),
self.num_atoms)
for i in range(len(self.dataframe)):
# ['acid_base', 'adjacency_inter', 'adjacency_intra', 'aromaticity', 'atomic_num', 'code', 'degree',
# 'distance', 'formal_charge', 'h_donor_acceptor', 'hybridization', 'implicit_valence', 'num_hs',
# 'output', 'ring_size', 'symbol']
delta = self.num_atoms - len(self.dataframe.iat[i, 0])
for j in [1, 2, 7]:
self.dataframe.iat[i, j] = np.pad(self.dataframe.iat[i, j],
((0, delta), (0, delta)),
'constant',
constant_values=((0, 0),
(0, 0)))
for j in [0, 3, 6, 8, 9, 10, 11, 12, 14, 15]:
self.dataframe.iat[i, j] = np.pad(
self.dataframe.iat[i, j], ((0, delta), (0, 0)),
'constant',
constant_values=((False, False), (False, False)))
self.dataframe.iat[i, 4] = np.pad(self.dataframe.iat[i, 4],
((0, delta), (0, 0)),
'constant',
constant_values=((0, 0), (0, 0)))
def geometric_matrix(mol, conformer=-1):
return rdmolops.Get3DDistanceMatrix(mol, confId=conformer)