scDeepJointClust: Deep neural network learning biological condition information refines gene-expression-based cell subtypes
The proposed method proceeds in four steps. Firstly, a DNN model is constructed and trained (see Figure. 1A below). Secondly, the representation data is retrieved from the trained DNN model and a clustering algorithm is performed on it (see Figure. 1B). After that, the clustering result is annotated using SingleR (16) (see Figure. 1C). Lastly, the method's performance is evaluated by considering the true positive rate and false positive rate (see Figure. 1D).
Figrue. 1. An overview of scDeepJointClust.
#Construct a model.
celltypeModel = buildModel(numInputCols,numOutputsCellType,numOutputsRorNR);
#Train the model.
celltypeModel = trainPhase1Model(celltypeModel,trainDataset,Configurations.numOfEpochs,Configurations.batchSize,Configurations.celltypeLossWeight, Configurations.RorNRLossWeight);
representationLayerModel = tf.keras.models.Model(inputs=celltypeModel.input,outputs=celltypeModel.get_layer("hiddenLayer3").output);
representationLayterOutput = representationLayerModel.predict(XTensor);
kmeans = KMeans(n_clusters=curK, random_state=0).fit(representationLayterOutput);
📧 Please let us know if you have any questions, bug reports, or feedback via the following email:
📧 hyp15@pitt.edu