DeepMol

Description

DeepMol is a Python-based machine and deep learning framework for drug discovery. It offers a variety of functionalities that enable a smoother approach to many drug discovery and chemoinformatics problems. It uses Tensorflow, Keras, Scikit-learn and DeepChem to build custom ML and DL models or make use of pre-built ones. It uses the RDKit framework to perform operations on molecular data.

More detailed and comprehensive documentation in DeepMol readthedocs.

Google colabs to run AutoML

• Binary and multiclass classification
• Regression
• Multi-task/multi-label

Installation

Pip

Install DeepMol via pip:

If you intend to install all the deepmol modules' dependencies:

pip install deepmol[all]

or in MacOS:

pip install "deepmol[all]"

Extra modules:

pip install deepmol[preprocessing]
pip install deepmol[machine-learning]
pip install deepmol[deep-learning]

or in MacOS:

pip install "deepmol[preprocessing]"
pip install "deepmol[machine-learning]"
pip install "deepmol[deep-learning]"

Also, you should install mol2vec and its dependencies:

pip install git+https://github.com/samoturk/mol2vec#egg=mol2vec

Manually

Alternatively, clone the repository and install the dependencies manually:

1. Clone the repository:

git clone https://github.com/BioSystemsUM/DeepMol.git

1. Install dependencies:

python setup.py install

Docker

You can also use the provided image to build your own Docker image:

docker pull biosystemsum/deepmol

Disclaimer

If you’d like to use the GPU, make sure to install the versions of TensorFlow and DGL that match the CUDA drivers for your hardware.

Do not install JAX, it will result dependency conflicts.

Loading tensorflow models will be problematic for MacOS users due to a known tensorflow issue 46.

Getting Started

DeepMol is built in a modular way allowing the use of its methods for multiple tasks. It offers a complete workflow to perform ML and DL tasks using molecules represented as SMILES. It has modules that perform standard tasks such as the loading and standardization of the data, computing molecular features like molecular fingerprints, performing feature selection and data splitting. It also provides methods to deal with unbalanced datasets, do unsupervised exploration of the data and compute feature importance as shap values.

Load a dataset from a CSV

To load data from a CSV it's only required to provide the math and molecules field name. Optionally, it is also possible to provide a field with some ids, the labels fields, features fields, features to keep (useful for instance to select only the features kept after feature selection) and the number of samples to load (by default loads the entire dataset).

from deepmol.loaders.loaders import CSVLoader

# load a dataset from a CSV (required fields: dataset_path and smiles_field)
loader = CSVLoader(dataset_path='../../data/train_dataset.csv',
                   smiles_field='mols',
                   id_field='ids',
                   labels_fields=['y'],
                   features_fields=['feat_1', 'feat_2', 'feat_3', 'feat_4'],
                   shard_size=1000,
                   mode='auto')

dataset = loader.create_dataset()

# print the shape of the dataset (molecules, X, y)
dataset.get_shape()

((1000,), None, (1000,))

Load a dataset from an SDF

If you want to load a dataset from an SDF file with 3D structures, it is only required to provide the path to the file. Optionally, it is also possible to provide a field with some ids, the labels fields.

from deepmol.loaders import SDFLoader

# load a dataset from a SDF (required fields: dataset_path)
loader = SDFLoader(dataset_path='../../data/train_dataset.sdf',
                   id_field='ids',
                   labels_fields=['y'],
                   shard_size=1000,
                   mode='auto')
dataset = loader.create_dataset()
dataset.get_shape()

((1000,), None, (1000,))

Compound Standardization

It is possible to standardize the loaded molecules using three options. Using a basic standardizer that only does sanitization (Kekulize, check valencies, set aromaticity, conjugation and hybridization). A more complex standardizer can be customized by choosing or not to perform specific tasks such as sanitization, removing isotope information, neutralizing charges, removing stereochemistry and removing smaller fragments. Another possibility is to use the ChEMBL Standardizer.

from deepmol.standardizer import BasicStandardizer, CustomStandardizer, ChEMBLStandardizer 

# Option 1: Basic Standardizer
standardizer = BasicStandardizer().standardize(dataset)

# Option 2: Custom Standardizer
heavy_standardisation = {
    'REMOVE_ISOTOPE': True,
    'NEUTRALISE_CHARGE': True,
    'REMOVE_STEREO': True,
    'KEEP_BIGGEST': True,
    'ADD_HYDROGEN': True,
    'KEKULIZE': False,
    'NEUTRALISE_CHARGE_LATE': True}
standardizer2 = CustomStandardizer(heavy_standardisation).standardize(dataset)

# Option 3: ChEMBL Standardizer
standardizer3 = ChEMBLStandardizer().standardize(dataset)

Compound Featurization

It is possible to compute multiple types of molecular fingerprints like Morgan Fingerprints, MACCS Keys, Layered Fingerprints, RDK Fingerprints and AtomPair Fingerprints. Featurizers from DeepChem and molecular embeddings like the Mol2Vec can also be computed. More complex molecular embeddings like the Seq2Seq and transformer-based are in development and will be added soon.

from deepmol.compound_featurization import MorganFingerprint

# Compute morgan fingerprints for molecules in the previously loaded dataset
MorganFingerprint(radius=2, size=1024).featurize(dataset, inplace=True)
# view the computed features (dataset.X)
dataset.X

#print shape of the dataset to see the difference in the X shape
dataset.get_shape()

((1000,), (1000, 1024), (1000,))

Feature Selection

Regarding feature selection it is possible to do Low Variance Feature Selection, KBest, Percentile, Recursive Feature Elimination and selecting features based on importance weights.

from deepmol.feature_selection import LowVarianceFS

# Feature Selection to remove features with low variance across molecules
LowVarianceFS(0.15).select_features(dataset, inplace=True)

# print shape of the dataset to see the difference in the X shape (fewer features)
dataset.get_shape()

((1000,), (1000, 35), (1000,))

Unsupervised Exploration

It is possible to do unsupervised exploration of the datasets using PCA, tSNE, KMeans and UMAP.

from deepmol.unsupervised import UMAP

ump = UMAP()
umap_df = ump.run(dataset)
ump.plot(umap_df.X, path='umap_output.png')

Data Split

Data can be split randomly or using stratified splitters. K-fold split, train-test split and train-validation-test split can be used.

from deepmol.splitters.splitters import SingletaskStratifiedSplitter

# Data Split
splitter = SingletaskStratifiedSplitter()
train_dataset, valid_dataset, test_dataset = splitter.train_valid_test_split(dataset=dataset, frac_train=0.7,
                                                                             frac_valid=0.15, frac_test=0.15)
train_dataset.get_shape()

((1628,), (1628, 1024), (1628,))

valid_dataset.get_shape()

((348,), (348, 1024), (348,))

test_dataset.get_shape()

((350,), (350, 1024), (350,))

Build, train and evaluate a model

It is possible to use pre-built models from Scikit-Learn and DeepChem or build new ones using keras layers. Wrappers for Scikit-Learn, Keras and DeepChem were implemented allowing evaluation of the models under a common workspace.

Scikit-Learn model example

Models can be imported from scikit-learn and wrapped using the SKlearnModel module.

from sklearn.ensemble import RandomForestClassifier
from deepmol.models.sklearn_models import SklearnModel

# Scikit-Learn Random Forest
rf = RandomForestClassifier()
# wrapper around scikit learn models
model = SklearnModel(model=rf)
# model training
model.fit(train_dataset)

from deepmol.metrics.metrics import Metric
from deepmol.metrics.metrics_functions import roc_auc_score

# cross validate model on the full dataset
best_model, train_score_best_model, test_score_best_model, \
            train_scores, test_scores, average_train_score, average_test_score = model.cross_validate(dataset, Metric(roc_auc_score), folds=3)

from sklearn.metrics import precision_score, accuracy_score, confusion_matrix, classification_report

#evaluate the model using different metrics
metrics = [Metric(roc_auc_score), Metric(precision_score), Metric(accuracy_score), Metric(confusion_matrix), 
           Metric(classification_report)]

# evaluate the model on training data
print('Training Dataset: ')
train_score = model.evaluate(train_dataset, metrics)

# evaluate the model on training data
print('Validation Dataset: ')
valid_score = model.evaluate(valid_dataset, metrics)

# evaluate the model on training data
print('Test Dataset: ')
test_score = model.evaluate(test_dataset, metrics)
model.save('my_model')

Loading and saving models was never so easy!

model = SklearnModel.load('my_model')
model.evaluate(test_dataset, metrics)

Keras model example

Example of how to build and wrap a keras model using the KerasModel module.

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from deepmol.metrics.metrics import Metric

input_dim = train_dataset.X.shape[1]


def create_model(optimizer='adam', dropout=0.5, input_dim=input_dim):
  # create model
  model = Sequential()
  model.add(Dense(12, input_dim=input_dim, activation='relu'))
  model.add(Dropout(dropout))
  model.add(Dense(8, activation='relu'))
  model.add(Dense(1, activation='sigmoid'))
  # Compile model
  model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
  return model


from deepmol.models.keras_models import KerasModel

model = KerasModel(create_model, epochs=5, verbose=1, optimizer='adam')

# train model
model.fit(train_dataset)

# make prediction on the test dataset with the model
model.predict(test_dataset)

# evaluate model using multiple metrics
metrics = [Metric(roc_auc_score),
           Metric(precision_score),
           Metric(accuracy_score),
           Metric(confusion_matrix),
           Metric(classification_report)]

print('Training set score:', model.evaluate(train_dataset, metrics))
print('Test set score:', model.evaluate(test_dataset, metrics))

model.save('my_model')

Loading and saving models was never so easy!

model = KerasModel.load('my_model')
model.evaluate(test_dataset, metrics)

DeepChem model example

Using DeepChem models:

from deepmol.compound_featurization import WeaveFeat
from deepchem.models import MPNNModel
from deepmol.models.deepchem_models import DeepChemModel
from deepmol.metrics.metrics import Metric
from deepmol.splitters.splitters import SingletaskStratifiedSplitter

ds = WeaveFeat().featurize(dataset)
splitter = SingletaskStratifiedSplitter()
train_dataset, valid_dataset, test_dataset = splitter.train_valid_test_split(dataset=ds, frac_train=0.6, frac_valid=0.2,
                                                                             frac_test=0.2)
mpnn = MPNNModel
model_mpnn = DeepChemModel(mpnn, n_tasks=1, n_pair_feat=14, n_atom_feat=75, n_hidden=75, T=1, M=1, mode='classification')
# Model training
model_mpnn.fit(train_dataset)
valid_preds = model_mpnn.predict(valid_dataset)
test_preds = model_mpnn.predict(test_dataset)
# Evaluation
metrics = [Metric(roc_auc_score), Metric(precision_score), Metric(accuracy_score)]
print('Training Dataset: ')
train_score = model_mpnn.evaluate(train_dataset, metrics)
print('Valid Dataset: ')
valid_score = model_mpnn.evaluate(valid_dataset, metrics)
print('Test Dataset: ')
test_score = model_mpnn.evaluate(test_dataset, metrics)
model_mpnn.save("my_model")

Loading and saving models was never so easy!

model = DeepChemModel.load('my_model')
model.evaluate(test_dataset, metrics)

Hyperparameter Optimization

Grid and randomized hyperparameter optimization is provided using cross-validation or a held-out validation set.

```python from deepmol.parameter_optimization.hyperparameter_optimization import HyperparameterOptimizerValidation

from tensorflow.keras.optimizers import Adam from tensorflow.keras.layers import Dropout from tensorflow import keras from tensorflow.keras import layers

def create_model(input_dim, optimizer='adam', dropout=0.5): # create model inputs = layers.Input(shape=input_dim)

# Define the shared layers
shared_layer_1 = layers.Dense(64, activation="relu")
dropout_1 = Dropout(dropout)
shared_layer_2 = layers.Dense(32, activation="relu")

# Define the shared layers for the inputs
x = shared_layer_1(inputs)
x = dropout_1(x)
x = shared_layer_2(x)

task_output = layers.Dense(1, activation="sigmoid")(x)

# Define the model that outputs the predictions for each task
model = keras.Model(inputs=inputs, outputs=task_output)
# Compile the model with different loss functions and metrics for each task
model.compile(
    optimizer=optimizer, loss="binary_crossentropy", metrics=["accuracy"]
)
return model

optimizer = HyperparameterOptimizerValidation(create_model,