The DeconvDLModel Class — DeconvDLModel-class • SpatialDDLS

The DeconvDLModel object stores all the information related to deep neural network models. It consists of the trained model, the training history, and the predictions on test data. After running calculateEvalMetrics, it is possible to find the performance evaluation of the model on test data (see ?calculateEvalMetrics for details).

Details

The steps related to Deep Learning are carried out using the keras and tensorflow packages, which use the R6 classes system. If you want to save the DeconvDLModel object as an RDS file, SpatialDDLS provides a saveRDS generic function that transforms the R6 object containing the trained model into a native valid R object. Specifically, the model is converted into a list with the architecture of the network and the weights learned during training, which is the minimum information needed to use the model as a predictor. If you want to keep the optimizer state, see ?saveTrainedModelAsH5. If you want to store either the DeconvDLModel or the SpatialDDLS objects on disk as RDA files, see ?preparingToSave.

Slots

model: Trained deep neural network. This slot can contain an R6 keras.engine.sequential.Sequential object or a list with two elements: the architecture of the model and the resulting weights after training.
training.history: List with the evolution of the selected metrics during training.
test.metrics: Performance of the model on test data.
test.pred: Predicted cell type proportions on test data.
cell.types: Vector with cell types considered by the model.
features: Vector with features (genes) considered by the model. These features will be used for subsequent predictions.
test.deconv.metrics: Performance of the model on test data by cell type. This slot is generated by the calculateEvalMetrics function (see ?calculateEvalMetrics for more details).
interpret.gradients: Gradients for interpretation. SpatialDDLS provides some functions to better understand prediction made by the model (see ?interGradientsDL for more details). This slot is a list of either one or two elements: gradients of either the loss function or the predicted class with respect to the input variables using pure (only one cell type) mixed transcriptional profiles. These gradients can be interpreted as to what extent the model is using these variables to predict each cell type proportions.