def create_RF_feature_extractors(extractor_kwargs,
n_estimators=[10, 100, 1000],
min_samples_leaves=[0.25, 0.1, 1],
max_depths=[None, 1, 10, 100]):
extractors = []
for one_vs_rest in [False, True]:
# We only consider min samples leafs and max_depths if we have real multiclass
for md in [None] if one_vs_rest else max_depths:
for msl in [1] if one_vs_rest else min_samples_leaves:
suffix = "" if one_vs_rest else "_multiclass"
if md is not None: # None is the default value
suffix += "_max_depth{}".format(msl)
if msl > 1: # 1 is the default vlaue
suffix += "_max_depth{}".format(msl)
for nest in n_estimators:
extractors.append(
fe.RandomForestFeatureExtractor(
name="{}-estimators{}".format(nest, suffix),
classifier_kwargs={
'n_estimators': nest,
'min_samples_leaf': msl,
'max_depth': md
},
one_vs_rest=one_vs_rest,
**extractor_kwargs))
return _shuffle_and_shorten(extractors)
logger = logging.getLogger("demo")
logger.setLevel('INFO')
# Create data for which we know the ground truth
dg = DataGenerator(natoms=20,
nclusters=2,
natoms_per_cluster=2,
nframes_per_cluster=500)
samples, labels = dg.generate_data()
feature_to_resids = dg.feature_to_resids
logger.info("Generated samples and labels of shapes %s and %s", samples.shape,
labels.shape)
# Identify important residues using a random forest
extractor = fe.RandomForestFeatureExtractor(samples=samples, labels=labels)
extractor = fe.PCAFeatureExtractor(
samples=samples) # Uncomment for unsupervised learning
extractor.extract_features()
# Postprocess the results to convert importance per feature into importance per residue
postprocessor = extractor.postprocessing()
postprocessor.average()
postprocessor.persist()
# Visualize the importance per residue
# Dashed lines show the residues we know are important (i.e. the atoms moved by the toy model)
visualization.visualize([[postprocessor]], highlighted_residues=dg.moved_atoms)
logger.info(
"Below we list all features and their importance. Those with highest importance are good candidates for CVs"
def run_toy_model(dg, data, labels, supervised=True, filetype="svg", n_iterations=10, variance_cutoff="1_components"):
cluster_indices = labels.argmax(axis=1)
feature_to_resids = dg.feature_to_resids
suffix = dg.test_model + "_" + dg.feature_type \
+ ("_supervised" if supervised else "_unsupervised") \
+ ("_var-cutoff=" + str(variance_cutoff) if not supervised else "")
kwargs = {
'samples': data,
'labels': cluster_indices,
'filter_by_distance_cutoff': False,
'use_inverse_distances': True,
'n_splits': 1,
'n_iterations': n_iterations,
# 'upper_bound_distance_cutoff': 1.,
# 'lower_bound_distance_cutoff': 1.
}
supervised_feature_extractors = [
fe.MlpFeatureExtractor(
activation="relu",
classifier_kwargs={
# 'hidden_layer_sizes': (dg.natoms, dg.nclusters * 2),
'hidden_layer_sizes': (int(dg.natoms / 2),),
# 'hidden_layer_sizes': [int(min(dg.nfeatures, 100) / (i + 1)) for i in range(10)],
'max_iter': 10000,
'alpha': 0.001,
},
per_frame_importance_outfile="output/toy_model_perframe.txt",
one_vs_rest=True,
**kwargs),
# fe.ElmFeatureExtractor(
# activation="relu",
# classifier_kwargs={
# 'hidden_layer_sizes': (dg.nfeatures,),
# 'alpha': 50,
# },
# **kwargs),
fe.KLFeatureExtractor(**kwargs),
fe.RandomForestFeatureExtractor(
one_vs_rest=True,
classifier_kwargs={'n_estimators': 100},
**kwargs),
]
unsupervised_feature_extractors = [
fe.MlpAeFeatureExtractor(
classifier_kwargs={
# hidden_layer_sizes=(int(data.shape[1]/2),),
'hidden_layer_sizes': (dg.nclusters,),
# 'hidden_layer_sizes': (10, 5, 1, 5, 10,),
# hidden_layer_sizes=(100, 1, 100,),
# hidden_layer_sizes=(200, 50, 10, 1, 10, 50, 200, ),
'max_iter': 100000,
# hidden_layer_sizes=(300, 200, 50, 10, 1, 10, 50, 200, 300,),
# max_iter=10000,
# 'alpha': 0.0001,
'alpha': 1,
'solver': "adam",
},
use_reconstruction_for_lrp=True,
activation="logistic",
**kwargs),
fe.PCAFeatureExtractor(classifier_kwargs={'n_components': None},
variance_cutoff=variance_cutoff,
name='PCA',
**kwargs),
# fe.RbmFeatureExtractor(classifier_kwargs={'n_components': dg.nclusters},
# relevance_method='from_lrp',
# name='RBM',
# **kwargs),
]
feature_extractors = supervised_feature_extractors if supervised else unsupervised_feature_extractors
logger.info("Done. using %s feature extractors", len(feature_extractors))
postprocessors = []
filter_results = False
for extractor in feature_extractors:
extractor.error_limit = 50
logger.info("Computing relevance for extractors %s", extractor.name)
extractor.extract_features()
p = extractor.postprocessing(working_dir="./{}".format(extractor.name),
pdb_file=None,
feature_to_resids=feature_to_resids,
filter_results=filter_results)
p.average()
p.evaluate_performance()
p.persist()
postprocessors.append([p])
logger.info("Done")
logger.info(
"Actual atoms moved: %s.\n(Cluster generation method %s. Noise level=%s, displacement=%s. frames/cluster=%s)",
sorted(dg.moved_atoms),
dg.test_model, dg.noise_level, dg.displacement, dg.nframes_per_cluster)
visualization.visualize(postprocessors,
show_importance=True,
show_performance=False,
show_projected_data=False,
highlighted_residues=dg.moved_atoms,
outfile="output/test_importance_per_residue_{suffix}.{filetype}".format(suffix=suffix,
filetype=filetype))
# visualization.visualize(postprocessors,
# show_importance=False,
# show_performance=True,
# show_projected_data=False,
# outfile="output/test_performance_{suffix}.{filetype}".format(suffix=suffix,
# filetype=filetype))
# visualization.visualize(postprocessors,
# show_importance=False,
# show_performance=False,
# show_projected_data=True,
# outfile="output/test_projection_{suffix}.{filetype}".format(suffix=suffix,
# filetype=filetype))
logger.info("Done. The settings were n_iterations = {n_iterations}, n_splits = {n_splits}."
"\nFiltering (filter_by_distance_cutoff={filter_by_distance_cutoff})".format(**kwargs))
def run_VSD(working_dir="bio_input/VSD/",
cluster_for_prediction=None,
dt_for_prediction=10,
multiclass=False):
data = np.load(working_dir + 'frame_i_j_contacts_dt1.npy')
cluster_indices = np.loadtxt(working_dir + 'clusters_indices.dat')
kwargs = {
'samples': data,
'labels': cluster_indices,
'filter_by_distance_cutoff': True,
'use_inverse_distances': True,
'n_splits': 3,
'n_iterations': 5,
'scaling': True,
'shuffle_datasets': True
}
if cluster_for_prediction is not None:
cluster_traj = md.load("{}/{}_dt{}.xtc".format(working_dir,
cluster_for_prediction,
dt_for_prediction),
top=working_dir + "alpha.pdb")
other_samples, _, _ = tp.to_distances(traj=cluster_traj,
scheme="closest-heavy",
pairs="all-residues",
use_inverse_distances=True,
ignore_nonprotein=True,
periodic=True)
logger.debug(
"Loaded cluster samples for prediction of shape %s for state %s",
other_samples.shape, cluster_for_prediction)
cluster_traj = None # free memory
else:
other_samples = False
feature_extractors = [
fe.RandomForestFeatureExtractor(
classifier_kwargs={'n_estimators': 100},
one_vs_rest=not multiclass,
**kwargs),
fe.KLFeatureExtractor(bin_width=0.1, **kwargs),
fe.MlpFeatureExtractor(
classifier_kwargs={
'hidden_layer_sizes': [
100,
],
'max_iter': 100000,
'alpha': 0.0001
},
activation="relu",
one_vs_rest=not multiclass,
per_frame_importance_samples=other_samples,
per_frame_importance_labels=
None, # If None the method will use predicted labels for LRP
per_frame_importance_outfile="{}/mlp_perframe_importance_{}/"
"VSD_mlp_perframeimportance_{}_dt{}.txt".format(
working_dir, "multiclass" if multiclass else "binaryclass",
cluster_for_prediction, dt_for_prediction),
**kwargs)
]
common_peaks = {
"R1-R4": [294, 297, 300, 303],
"K5": [306],
"R6": [309],
}
do_computations = True
filetype = "svg"
for extractor in feature_extractors:
logger.info("Computing relevance for extractors %s", extractor.name)
extractor.extract_features()
p = extractor.postprocessing(working_dir=working_dir,
pdb_file=working_dir + "alpha.pdb",
filter_results=False)
if do_computations:
p.average()
p.evaluate_performance()
p.persist()
else:
p.load()
visualization.visualize(
[[p]],
show_importance=True,
show_performance=False,
show_projected_data=False,
highlighted_residues=common_peaks,
outfile=working_dir +
"{extractor}/importance_per_residue_{suffix}.{filetype}".format(
suffix="", extractor=extractor.name, filetype=filetype))
if do_computations:
visualization.visualize(
[[p]],
show_importance=False,
show_performance=True,
show_projected_data=False,
outfile=working_dir +
"{extractor}/performance_{suffix}.{filetype}".format(
extractor=extractor.name, suffix="", filetype=filetype))
visualization.visualize(
[[p]],
show_importance=False,
show_performance=False,
show_projected_data=True,
outfile=working_dir +
"{extractor}/projected_data_{suffix}.{filetype}".format(
extractor=extractor.name, suffix="", filetype=filetype))
logger.info("Done")
def run_CaM(parser):
# Known important residues
common_peaks = [109, 144, 124, 145, 128, 105, 112, 136, 108, 141, 92]
shuffle_data = True
args = parser.parse_args()
working_dir = args.out_directory
n_runs = args.number_of_runs
samples = np.load(args.feature_list)
cluster_indices = np.loadtxt(args.cluster_indices)
# Shift cluster indices to start at 0
cluster_indices -= cluster_indices.min()
if shuffle_data:
# Permute blocks of 100 frames
n_samples = samples.shape[0]
n_samples = int(n_samples / 100) * 100
inds = np.arange(n_samples)
inds = inds.reshape((int(n_samples / 100), 100))
perm_inds = np.random.permutation(inds)
perm_inds = np.ravel(perm_inds)
samples = samples[perm_inds]
cluster_indices = cluster_indices[perm_inds]
pdb_file = args.pdb_file
labels = cluster_indices
lower_distance_cutoff = 1.0
upper_distance_cutoff = 1.0
n_components = 20
# Check if samples format is correct
if len(samples.shape) != 2:
sys.exit("Matrix with features should have 2 dimensions")
kwargs = {
'samples': samples,
'labels': labels,
'filter_by_distance_cutoff': True,
'lower_bound_distance_cutoff': lower_distance_cutoff,
'upper_bound_distance_cutoff': upper_distance_cutoff,
'use_inverse_distances': True,
'n_splits': args.number_of_k_splits,
'n_iterations': args.number_of_iterations,
'scaling': True
}
feature_extractors = [
fe.PCAFeatureExtractor(variance_cutoff=0.75, **kwargs),
fe.RbmFeatureExtractor(relevance_method="from_components", **kwargs),
fe.MlpAeFeatureExtractor(activation=relprop.relu,
classifier_kwargs={
'solver': 'adam',
'hidden_layer_sizes': (100, )
},
**kwargs),
fe.RandomForestFeatureExtractor(
one_vs_rest=True,
classifier_kwargs={'n_estimators': 500},
**kwargs),
fe.KLFeatureExtractor(**kwargs),
fe.MlpFeatureExtractor(classifier_kwargs={
'hidden_layer_sizes': (120, ),
'solver': 'adam',
'max_iter': 1000000
},
activation=relprop.relu,
**kwargs),
]
postprocessors = []
for extractor in feature_extractors:
tmp_pp = []
for i_run in range(n_runs):
extractor.extract_features()
# Post-process data (rescale and filter feature importances)
p = extractor.postprocessing(working_dir=working_dir,
rescale_results=True,
filter_results=False,
feature_to_resids=None,
pdb_file=pdb_file)
p.average().evaluate_performance()
p.persist()
# Add common peaks
tmp_pp.append(p)
postprocessors.append(tmp_pp)
visualization.visualize(
postprocessors,
show_importance=True,
show_projected_data=False,
show_performance=False,
highlighted_residues=common_peaks,
outfile="{}/importance-per-residue.png".format(working_dir))
logger.info("Done")
def run_beta2(
working_dir="bio_input/beta2/",
n_iterations=1,
n_splits=1,
shuffle_datasets=True,
overwrite=False,
dt=1,
feature_type="ca_inv", # "closest-heavy_inv", "CA_inv", "cartesian_ca", "cartesian_noh" or "compact_ca_inv"
filetype="svg",
classtype="multiclass",
supervised=True,
load_trajectory_for_predictions=False,
filter_by_distance_cutoff=False,
ligand_type='holo'):
results_dir = "{}/results/{}/{}/{}/".format(
working_dir, classtype, feature_type,
"cutoff" if filter_by_distance_cutoff else "nocutoff")
samples_dir = "{}/samples/{}/{}".format(working_dir, classtype,
feature_type)
data = np.load("{}/samples_dt{}.npz".format(samples_dir, dt))['array']
feature_to_resids = np.load("{}/feature_to_resids.npy".format(
samples_dir, feature_type))
labels = np.loadtxt(
"{wd}/cluster_indices/{ct}/cluster_indices_dt{dt}.txt".format(
wd=working_dir, ct=classtype, dt=dt))
if classtype == "multiclass":
label_names = ["agonist-bound", "protonated-asp79"]
mixed_classes = True
else:
label_names = ["apo", "holo"]
mixed_classes = False
suffix = str(-1) + "clusters_" + str(n_iterations) + "iterations_" \
+ ("distance-cutoff_" if filter_by_distance_cutoff else "") + feature_type
labels -= labels.min()
if len(data) != len(labels) or data.shape[1] != len(feature_to_resids):
raise Exception(
"Inconsistent input data. The number of features or the number of frames to no match"
)
logger.info("Loaded data of shape %s for feature type %s", data.shape,
feature_type)
# ## Define the different methods to use
# Every method is encapsulated in a so called FeatureExtractor class which all follow the same interface
cutoff_offset = 0.2 if "closest-heavy" in feature_type else 0
kwargs = {
'samples':
data,
'labels':
labels,
'label_names':
label_names,
'filter_by_distance_cutoff':
filter_by_distance_cutoff,
'lower_bound_distance_cutoff':
filtering.lower_bound_distance_cutoff_default - cutoff_offset,
'upper_bound_distance_cutoff':
filtering.upper_bound_distance_cutoff_default - cutoff_offset,
'use_inverse_distances':
True,
'n_splits':
n_splits,
'n_iterations':
n_iterations,
'shuffle_datasets':
shuffle_datasets
# 'upper_bound_distance_cutoff': 1.,
# 'lower_bound_distance_cutoff': 1.
}
unsupervised_feature_extractors = [
fe.PCAFeatureExtractor(
classifier_kwargs={'n_components': None},
variance_cutoff='auto',
# variance_cutoff='1_components',
name='PCA',
**kwargs),
fe.RbmFeatureExtractor(classifier_kwargs={'n_components': 1},
relevance_method='from_lrp',
name='RBM',
**kwargs),
# fe.MlpAeFeatureExtractor(
# classifier_kwargs={
# 'hidden_layer_sizes': (100, 30, 2, 30, 100,), # int(data.shape[1]/2),),
# # max_iter=10000,
# 'alpha': 0.01,
# 'activation': "logistic"
# },
# use_reconstruction_for_lrp=True,
# **kwargs),
]
if load_trajectory_for_predictions:
other_samples, other_labels = _load_trajectory_for_predictions(
ligand_type)
else:
other_samples, other_labels = None, None
supervised_feature_extractors = [
# fe.ElmFeatureExtractor(
# activation="relu",
# n_nodes=data.shape[1] * 2,
# alpha=0.1,
# **kwargs),
fe.KLFeatureExtractor(**kwargs),
fe.RandomForestFeatureExtractor(
one_vs_rest=True,
classifier_kwargs={'n_estimators': 500},
**kwargs),
fe.MlpFeatureExtractor(
name="MLP" if other_samples is None else
"MLP_predictor_{}".format(ligand_type),
classifier_kwargs={
# 'hidden_layer_sizes': [int(min(100, data.shape[1]) / (i + 1)) + 1 for i in range(3)],
'hidden_layer_sizes': (30, ),
# 'max_iter': 10000,
'alpha': 0.1,
'activation': "relu"
},
# per_frame_importance_samples=other_samples,
# per_frame_importance_labels=other_labels,
# per_frame_importance_outfile="/home/oliverfl/projects/gpcr/mega/Result_Data/beta2-dror/apo-holo/trajectories"
# "/mlp_perframe_importance_{}/"
# "{}_mlp_perframeimportance_{}clusters_{}cutoff.txt"
# .format(ligand_type, feature_type, nclusters, "" if filter_by_distance_cutoff else "no"),
**kwargs),
]
if supervised is None:
feature_extractors = unsupervised_feature_extractors + supervised_feature_extractors
else:
feature_extractors = supervised_feature_extractors if supervised else unsupervised_feature_extractors
logger.info("Done. using %s feature extractors", len(feature_extractors))
highlighted_residues = _get_important_residues(supervised, feature_type)
# # Run the relevance analysis
postprocessors = []
for extractor in feature_extractors:
do_computations = True
if os.path.exists(results_dir):
existing_files = glob.glob(
"{}/{}/importance_per_residue.npy".format(
results_dir, extractor.name))
if len(existing_files) > 0 and not overwrite:
logger.debug("File %s already exists. skipping computations",
existing_files[0])
do_computations = False
if do_computations:
logger.info("Computing relevance for extractors %s",
extractor.name)
extractor.extract_features()
p = extractor.postprocessing(
working_dir=results_dir,
pdb_file=working_dir + "/trajectories/all.pdb",
# pdb_file=working_dir + "/trajectories/protein_noh.pdb",
feature_to_resids=feature_to_resids,
filter_results=False)
if do_computations:
p.average()
p.evaluate_performance()
p.persist()
else:
p.load()
postprocessors.append([p])
# # Visualize results
visualization.visualize(
[[p]],
show_importance=True,
show_performance=False,
show_projected_data=False,
mixed_classes=mixed_classes,
highlighted_residues=highlighted_residues,
outfile=results_dir +
"/{extractor}/importance_per_residue_{suffix}_{extractor}.{filetype}"
.format(suffix=suffix, extractor=extractor.name,
filetype=filetype))
if do_computations:
visualization.visualize(
[[p]],
show_importance=False,
show_performance=True,
show_projected_data=False,
mixed_classes=mixed_classes,
outfile=results_dir +
"/{extractor}/performance_{suffix}_{extractor}.{filetype}".
format(suffix=suffix,
extractor=extractor.name,
filetype=filetype))
visualization.visualize(
[[p]],
show_importance=False,
show_performance=False,
show_projected_data=True,
mixed_classes=mixed_classes,
outfile=results_dir +
"/{extractor}/projected_data_{suffix}_{extractor}.{filetype}".
format(suffix=suffix,
extractor=extractor.name,
filetype=filetype))
logger.info(
"Done. The settings were n_iterations = {n_iterations}, n_splits = {n_splits}."
"\nFiltering (filter_by_distance_cutoff={filter_by_distance_cutoff})".
format(**kwargs))