kinoml.ml.torch_models

Module Contents

class kinoml.ml.torch_models._BaseModule

Bases: torch.nn.Module

needs_input_shape = True
static estimate_input_shape(input_sample)

This static method takes the same input as .forward() would and estimates the incoming shape so the layers can be initialized properly.

Most of the time, input_sample would be a Tensor, in which the first dimension corresponds to systems, and the second is the input shape we need.

If your .forward() method takes something else than a Tensor, please adapt this method accordingly.

class kinoml.ml.torch_models.NeuralNetworkRegression(input_shape, hidden_shape=100, output_shape=1, activation=F.relu)

Bases: _BaseModule

Builds a vanilla neural network and a feed-forward pass for a ligand setting.

Parameters

  • input_shape (int) – Dimension of the input vector.

  • hidden_shape (int, default=100) – Number of units in the hidden layer.

  • output_shape (int, default=1) – Size of the last unit, representing delta_g_over_kt in our setting.

  • _activation (torch function, default: relu) – The activation function used in the hidden (only!) layer of the network.

forward(system)

Defines the foward pass for a system, which contains the featurized molecular components and the associated measurement/label. Since only the featurized system is needed, the first element is retrieved. In this case, the ligand is a 1D array of shape (length of fingerprint).

Notes

More information about how featurization are stored can be found at https://github.com/openkinome/experiments-binding-affinity/blob/master/features/featurize-template.ipynb .

class kinoml.ml.torch_models.ListOfTupleNeuralNetworkregression(input_shape, hidden_shape=100, output_shape=1, activation=F.relu)

Bases: NeuralNetworkRegression

This example model does not take a Tensor in, but a tuple of tensors. Each tensor has shape (n_systems, n_features).

As a result, one needs to concatenate the results before passing it to the parent .forward() method.

static estimate_input_shape(input_sample)

This static method takes the same input as .forward() would and estimates the incoming shape so the layers can be initialized properly.

Most of the time, input_sample would be a Tensor, in which the first dimension corresponds to systems, and the second is the input shape we need.

If your .forward() method takes something else than a Tensor, please adapt this method accordingly.

forward(x)

Defines the foward pass for a system, which contains the featurized molecular components and the associated measurement/label. Since only the featurized system is needed, the first element is retrieved. In this case, the ligand is a 1D array of shape (length of fingerprint).

Notes

More information about how featurization are stored can be found at https://github.com/openkinome/experiments-binding-affinity/blob/master/features/featurize-template.ipynb .

class kinoml.ml.torch_models.DenseNeuralNetworkRegression(input_shape, hidden_shape=(350, 200, 100, 50, 16), output_shape=1, dropout_percentage=0.4, activation=F.relu)

Bases: _BaseModule

Builds a Dense Neural Network and a feed-forward pass.

Parameters

  • input_shape (int) – Dimension of the input vector.

  • hidden_shape (list) – Number of units in each of the hidden layers.

  • output_shape (int, default=1) – Size of the last unit, representing delta_g_over_kt in our setting.

  • dropout_percentage (float) – The percentage of hidden to by dropped at random.

  • _activation (torch function, default=relu) – The activation function used in the hidden (only!) layer of the network.

forward(x)

Defines the foward pass for a given input ‘x’

class kinoml.ml.torch_models.ConvolutionNeuralNetworkRegression(input_shape, embedding_shape=300, kernel_shape=10, hidden_shape=100, output_shape=1, activation=F.relu)

Bases: _BaseModule

Builds a Convolutional Neural Network and a feed-forward pass for a ligand setting.

Parameters

  • input_shape (int) – Dimension of the input of the system.

  • embedding_shape (int, default=200) – Dimension of the embedding after convolution.

  • kernel_shape (int, default=10) – Size of the kernel for the convolution.

  • hidden_shape (int, default=100) – Number of units in the hidden layer.

  • output_shape (int, default=1) – Size of the last unit, representing delta_g_over_kt in our setting.

  • activation (torch function, default=relu) – The activation function used in the hidden (only!) layer of the network.

forward(system)

Defines the foward pass for a system, which contains the featurized molecular components and the associated measurement/label. Since only the featurized system is needed, the first element is retrieved. In this case, the ligand is a 2D array of shape (nb of character, length of smiles).

Notes

More information about how featurization are stored can be found at https://github.com/openkinome/experiments-binding-affinity/blob/master/features/featurize-template.ipynb .

class kinoml.ml.torch_models.ConvolutionNeuralNetworkRegressionProteinInformed(input_shape, embedding_shape=300, kernel_shape=10, hidden_shape=100, output_shape=1, activation=F.relu)

Bases: _BaseModule

Builds a Convolutional Neural Network and a feed-forward pass for a protein-ligand setting.

Parameters

  • input_shape (tuple) – Dimension of input tensors, for the ligand and the protein.

  • embedding_shape (int, default=200) – Dimension of the embedding after convolution.

  • kernel_shape (int, default=10) – Size of the kernel for the convolution.

  • hidden_shape (int, default=100) – Number of units in the hidden layer.

  • output_shape (int, default=1) – Size of the last unit, representing delta_g_over_kt in our setting.

  • activation (torch function, default=relu) – The activation function used in the hidden (only!) layer of the network.

forward(system)

Defines the foward pass for given a system, which contains the featurized molecular components and the associated measurement/label. In this case, the input is composed of two entities: ligand and protein, so the two first elements are retrieved. In this case, the ligand is a 2D array of shape (nb of character, length of smiles) and the protein a 2D array of shape (nb of residues, length of sequence)

Notes

More information about how featurization are stored can be found at https://github.com/openkinome/experiments-binding-affinity/blob/master/features/featurize-template.ipynb .

class kinoml.ml.torch_models.NeuralNetworkRegressionProteinInformed(input_shape, embedding_shape_ligand=300, embedding_shape_protein=10, kernel_shape=10, hidden_shape=100, output_shape=1, activation=F.relu)

Bases: _BaseModule

Builds a Neural Network and a feed-forward pass for a protein-ligand setting.

Parameters

  • input_shape (tuple) – Dimension of input tensors, for the ligand and the protein.

  • embedding_shape_ligand (int, default=300) – Dimension of the embedding for the ligand.

  • embedding_shape_protein (int, default=10) – Dimension of the embedding for the protein.

  • kernel_shape (int, default=10) – Size of the kernel for the convolution.

  • hidden_shape (int, default=100) – Number of units in the hidden layer.

  • output_shape (int, default=1) – Size of the last unit, representing delta_g_over_kt in our setting.

  • activation (torch function, default=relu) – The activation function used in the hidden (only!) layer of the network.

forward(system)

Defines the foward pass for given a system, which contains the featurized molecular components and the associated measurement/label. In this case, the input is composed of two entities: ligand and protein, so the two first elements are retrieved. In this case, the ligand is a 1D array of shape (length of fingerprint), and the protein a 1D array of shape (length of sequence composition).

Notes

More information about how featurization are stored can be found at https://github.com/openkinome/experiments-binding-affinity/blob/master/features/featurize-template.ipynb .