cellmaps_coembedding.proteinprojector package

Submodules

cellmaps_coembedding.proteinprojector.architecture module

class cellmaps_coembedding.proteinprojector.architecture.Modality(training_data, name, transform, device)

Bases: object

Represents a single modality of data, containing training features and labels.

Initialize the Modality object with given training data, a name, a transformation, and the device.

Parameters:

• training_data – The data to use for training. Expects a list of lists where each sublist contains the label followed by feature values.

• name – The name of the modality.

• transform – The transformation to apply to the data, converting it to a tensor.

• device – The device to transfer the tensors to.

class cellmaps_coembedding.proteinprojector.architecture.Protein_Dataset(modalities_dict)

Bases: Dataset

A dataset class for handling protein data across multiple modalities.

Initialize the dataset using a dictionary of modalities.

Parameters:

modalities_dict – A dictionary where keys are modality names and values are Modality objects.

class cellmaps_coembedding.proteinprojector.architecture.ToTensor

Bases: object

A class that converts a numpy ndarray to a torch tensor.

class cellmaps_coembedding.proteinprojector.architecture.TrainingDataWrapper(modality_data, modality_names, device, l2_norm, dropout, latent_dim, hidden_size_1, hidden_size_2, resultsdir)

Bases: object

Wraps training data for all modalities.

Initialize the wrapper with the given configuration.

cellmaps_coembedding.proteinprojector.architecture.init_weights(module)

Initialize weights for linear layers using Xavier normal distribution and biases to zero.

class cellmaps_coembedding.proteinprojector.architecture.uniembed_nn(data_wrapper)

Bases: Module

A neural network model for embedding proteins using multiple modalities.

Initialize the model using a data wrapper that contains modality data configurations.

forward(inputs)

Forward pass of the model, processing inputs through encoders and decoders.

Parameters:

inputs – Dictionary of inputs where keys are modality names and values are corresponding tensors.

Returns:

Tuple of dictionaries containing latent representations and outputs for all modalities.

Module contents

ProteinProjector co-embedding algorithm.

This module provides the ProteinProjector implementation, formerly known as ProteinGPS. It exposes utilities for training the neural network, saving results, and yielding co-embeddings.

class cellmaps_coembedding.proteinprojector.Modality(training_data, name, transform, device)

Bases: object

Represents a single modality of data, containing training features and labels.

Initialize the Modality object with given training data, a name, a transformation, and the device.

Parameters:

• training_data – The data to use for training. Expects a list of lists where each sublist contains the label followed by feature values.

• name – The name of the modality.

• transform – The transformation to apply to the data, converting it to a tensor.

• device – The device to transfer the tensors to.

class cellmaps_coembedding.proteinprojector.Protein_Dataset(modalities_dict)

Bases: Dataset

A dataset class for handling protein data across multiple modalities.

Initialize the dataset using a dictionary of modalities.

Parameters:

modalities_dict – A dictionary where keys are modality names and values are Modality objects.

class cellmaps_coembedding.proteinprojector.ToTensor

Bases: object

A class that converts a numpy ndarray to a torch tensor.

class cellmaps_coembedding.proteinprojector.TrainingDataWrapper(modality_data, modality_names, device, l2_norm, dropout, latent_dim, hidden_size_1, hidden_size_2, resultsdir)

Bases: object

Wraps training data for all modalities.

Initialize the wrapper with the given configuration.

cellmaps_coembedding.proteinprojector.fit_predict(resultsdir: str, modality_data: Iterable[Iterable[Iterable[float]]], modality_names: Iterable[str] = (), batch_size: int = 16, latent_dim: int = 128, n_epochs: int = 250, triplet_margin: float = 1.0, lambda_reconstruction: float = 1.0, lambda_triplet: float = 1.0, lambda_l2: float = 0.001, l2_norm: bool = False, dropout: float = 0.0, save_epoch: int = 50, learn_rate: float = 0.0001, hidden_size_1: int = 512, hidden_size_2: int = 256, save_update_epochs: bool = False, mean_losses: bool = False, negative_from_batch: bool = False) → Generator[List[float], None, None]

Trains and predicts using a deep learning model with the given configuration and data.

Parameters:

• resultsdir – Directory to save training results and models.

• modality_data – Input data for the model.

• modality_names – Names of modalities; autogenerated if not provided.

• batch_size – Batch size for training.

• latent_dim – Dimensionality of the latent embeddings.

• n_epochs – Number of training epochs.

• triplet_margin – Margin for triplet loss.

• lambda_reconstruction – Weight for reconstruction loss.

• lambda_triplet – Weight for triplet loss.

• lambda_l2 – Weight for L2 regularization.

• l2_norm – Whether to use L2 normalization.

• dropout – Dropout rate.

• save_epoch – Epoch interval at which to save the model.

• learn_rate – Learning rate for the optimizer.

• hidden_size_1 – Size of the first hidden layer.

• hidden_size_2 – Size of the second hidden layer.

• save_update_epochs – Flag to save model state at specified epoch intervals.

• mean_losses – Whether to average losses or not.

• negative_from_batch – Whether to use negative samples from the same batch for triplet loss.

Returns:

Generator of average embeddings for each protein.

cellmaps_coembedding.proteinprojector.save_results(model: Module, protein_dataset: Protein_Dataset, data_wrapper: TrainingDataWrapper, results_suffix: str = '') → Dict[str, Dict[str, ndarray]]

Evaluates the model, saves the state, and exports embeddings for each protein.

Parameters:

• model – The neural network model.

• protein_dataset – The dataset containing protein data.

• data_wrapper – Data handling and configurations as an object.

• results_suffix – Suffix to append to results directory for saving.

Returns:

Dictionary keyed by protein containing modality embeddings.

class cellmaps_coembedding.proteinprojector.uniembed_nn(data_wrapper)

Bases: Module

A neural network model for embedding proteins using multiple modalities.

Initialize the model using a data wrapper that contains modality data configurations.

forward(inputs)

Forward pass of the model, processing inputs through encoders and decoders.

Parameters:

inputs – Dictionary of inputs where keys are modality names and values are corresponding tensors.

Returns:

Tuple of dictionaries containing latent representations and outputs for all modalities.

cellmaps_coembedding.proteinprojector.write_embedding_dictionary_to_file(filepath: str, dictionary: Dict[str, ndarray], dims: int) → None

Writes a dictionary of embeddings to a tab-separated file with headers.

Parameters:

• filepath – Path to the file where embeddings will be saved.

• dictionary – Dictionary of embeddings with keys as names and values as embedding vectors.

• dims – Dimension of embedding vectors.