deepnog.models package

deepnog.models.deepencoding module

deepnog.models.deepfam module

Author: Lukas Gosch, Roman Feldbauer

class deepnog.models.deepfam.DeepFam(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

Convolutional network for protein family prediction.

PyTorch implementation of DeepFam architecture (original: TensorFlow).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) beeing in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation1(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation12(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation123(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) beeing in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation13(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation2(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation23(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) beeing in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

class deepnog.models.deepfam.DeepFamAblation3(*args: Any, **kwargs: Any)

Bases: deepnog.models.deepfam.DeepFamAblationBase

Ablation study of DeepFam to DeepNOG transition.

Change 1: WordEmbedding instead of PseudoOneHot Change 2: SELU instead of BN/ReLU Change 3: Drop the fully connected layer between ConvNet and classification Combinations: 12, 13, 23, 123 (=DeepNOG).

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)

deepnog.models.deepnog module

Author: Lukas Gosch, Roman Feldbauer

Date: 2019-10-09

Description: Convolutional networks for protein orthologous group assignment.

class deepnog.models.deepnog.AminoAcidWordEmbedding(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

PyTorch nn.Embedding where each amino acid is considered one word.

Parameters

embedding_dim (int) – Embedding dimensionality.

forward(sequence)

Embed a given sequence.

Parameters

sequence (Tensor) – The sequence or a batch of sequences to embed. They are assumed to be translated to numerical values given a generated vocabulary (see gen_amino_acid_vocab in dataset.py)

Returns

x – The sequence (densely) embedded in a space of dimension embedding_dim.

Return type

Tensor

class deepnog.models.deepnog.DeepNOG(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

Convolutional network for protein orthologous group prediction.

Compared to DeepFam, this architecture provides:

• learned amino acid embeddings

• self-normalizing network with SELU

• sequence length independence

• stream-lined output layer

This networks consists of an embedding layer which learns a D-dimensional embedding for each amino acid. For a sequence of length L, the embedding has dimension DxL. A 1-D convolution with C filters of F different kernel- sizes K_i are performed over the embedding resulting in Cx(L-K_i-1) output dimension for each kernel size. SeLU activation is applied on the output followed by AdaptiveMaxPooling1D Layer reducing the dimension to of the output layer to Cx1 and resulting in the NN being sequence length independent. The max-pooling layer is followed up by a classic dropout Layer and then by a dense layer with as many output nodes as orthologous groups/protein families to classify.

Parameters

model_dict (dict) – Dictionary storing the hyperparameters and learned parameters of the model.

Notes

This architecture’s working title was DeepEncoding. The old name was last available in deepnog 1.2.2.

forward(x)

Forward a batch of sequences through network.

Parameters

x (Tensor, shape (batch_size, sequence_len)) – Sequence or batch of sequences to classify. Assumes they are translated using a vocabulary. (See gen_amino_acid_vocab in dataset.py)

Returns

out – Confidence of sequence(s) being in one of the n_classes.

Return type

Tensor, shape (batch_size, n_classes)