pychemlp

Easy-to-use python module for the training of Multi-Layer Perceptrons (neural networks) from molecular SMILES data and known associated properties.

This code was developed while doing my post-doc in the Zwijnenburg group, https://www.zwijnenburg-group.org/.

Functionality

• Import data from external .csv files and pickled objects
• Generate molecular fingerprints based on the Extended Connectivity Fingerprint (ECFP) method
• Specify neural network architectures (number of layers, neurons per layer, dropout)
• Train and evaluate models
• Conduct hyperparameter optimization using a random search

Using the module

Simple construction and training of a neural network

from pychemlp import MLP

nn = MLP()
nn.load_data('training_data.csv', 'SMILES', ['PROP1', 'PROP2', 'PROP3'])
nn.fingerprint(bits=2048, rad=2, test_frac=0.3)
nn.build_network(2, 256, dropout=0.5, activation='relu', input_dropout=0.5)
nn.train(epochs=10, batch_size=50, loss='mean_absolute_error)
y, pred, mae, rmse = nn.evaluate()

Where 'SMILES' is the name of the data column containing SMILES data and ['PROP1', 'PROP2', 'PROP3'] is a list of columns which contain the training target data (in this case, the model will be trained to predict three different properties).

Hyperpearameter optimization using a random search

from pychemlp import MLP

nn = MLP()
nn.load_data('./example_data.pkl', 'SMILES', ['PROP1', 'PROP2', 'PROP3'], from_pkl=True)
nn.fingerprint(bits=2048, rad=2, test_frac=0.3)

search_space = {'dropout': [0.1, 0.2, 0.3, 0.4, 0.5],
                'input_dropout': [0.1, 0.2, 0.3, 0.4, 0.5],
                'n_layers': [1, 2, 3, 4],
                'n_neurons': [32, 64, 128, 256, 512],
                'learning_rate': [0.1, 0.01, 0.001],
                'batch_size': [32, 64, 128, 256]}
nn.hyperparam_opt_random(search_space, 20, epochs=20)

Using a trained model

Once trained a model can be called as an attribute of the MLP class to make new predictions:

predictions = nn.model.predict(fingerprint_array)

Where fingerprint_array is an array of (Morgan) fingerprints which can be obtained by calling load_data and fingerprint, using a file containing SMILES for which predictions are to be made.

Loading and saving models

Existing models can be loaded or trained model saved

nn.load_model(path)
nn.save_model(path)

Where path is the path and filename of the save/load location. Models should be saved to/loaded form HDF5 format.

Installation & Requirements

To install, simply clone the repository:

git clone https://github.com/LiamWilbraham/pychemlp.git

and add the location of the pychemlp repository to your PYTHONPATH.

The module requires the following packages (all can be installed via conda):

• tensorflow (https://www.tensorflow.org/install/)
• scikit-learn (https://scikit-learn.org/stable/install.html)
• pandas (https://anaconda.org/anaconda/pandas)
• numpy (https://anaconda.org/anaconda/numpy)
• rdkit (https://www.rdkit.org/docs/Install.html)

rdkit may be installed via conda (recommended):

conda install -c rdkit rdkit