scDAC
scDAC is a single-cell Deep Adaptive Clustering (scDAC) model by coupling the Autoencoder and the Dirichlet Process Mixture Model.
Python --- 3.7.11
Pytorch --- 1.12.0
Sklearn --- 1.0.2
Numpy --- 1.21.6
Take the dataset chen_10 we provided here as an example.
Decompress the expression matrix file in scDAC/data/mat and split it for the input by running the scDAC/preprocess/preprocess_split.ipynb and obtain the files of cell names and feature names by running scDAC/preprocess/preprocess_cell_and_feature_names.ipynb
After running the preprocess_split or decompressing files vec.rar in the file path of scDAC/data/chen_10/subset_0 , run the following command to train the scDAC model:
$ CUDA_VISIBLE_DEVICES=0 py run.py --task chen_10 --exp e0
To infer the labels after training, run the following command:
$ CUDA_VISIBLE_DEVICES=0 py run.py --action infer_latent --task chen_10 --exp e0
You can obtain the predicted clustering labels under the folder /data/chen_10, and the ARI, NMI and SC metrics. The indicator of Deviation Ratio (DR) can be calculated by running "scDAC/metrics_DR.ipynb".
The results of scDAC in the paper can be reproduced by using the code in the file "scDAC/demo". The code of compared methods can be found in the file "scDAC/compared methods".
The datasets used in the paper can be found at (https://drive.google.com/file/d/1qyaxzW7k0rA9Y08P_ZlapVb43l0xZuoE/view?usp=drive_link), we can download and unzip them to the path "/data/preprocessed". When using the datasets orozco and kozareva, we can use "/preprocess/preprocess_split.ipynb" to obtain the input as the input of these two dataset are too large to upload. We can also download the original data and preprocess it as the following preprocess steps suggested in the following. The accession IDs are provided in the paper.
If you want to use scDAC to analyse other data, you can running the code as the following:
1.Please preprocess the data by Seurat referring to https://satijalab.org/seurat/articles/pbmc3k_tutorial, or use the code "preprocess.ipynb" provided by scDAC in the "preprocess" folder.
For data preprocessing, there are no specific requirements. Common preprocessing methods for scRNA-seq data, such as Seurat, Scanpy, and scran, are all acceptable. The general preprocessing steps involve normalizing and log-transforming the UMI count matrices. Afterwards, the final step is to select 4000 highly variable genes and scale them to obtain the input expression matrices. When researchers apply scDAC to large-scale scRNA-seq datasets, they can consider retaining a larger number of high-variable genes. We need to get the following files:
a. The expression matrices of highly variable genes.
b.The feature number and feature name of the highly variable genes and the cellnames of every expression matrix.
We can use "preprocess_split.ipynb" in the "preprocess" folder to split the matrix to vectors. The vectors are the inneed input of scDAC.
2.Please supplement the file "scDAC/configs/data.toml" with information about the new dataset. For parameter tuning, users can explore different parameters in the AE module to accommodate changes in the dataset scale. The default parameters in the AE module are as follows: the encoder consists of two hidden layers with sizes of 256 and 128, while the decoder has hidden layers with sizes of 128 and 256. The low-dimensional representation is set to a dimensionality of 32. During the training process, the minibatch size is set to 512. When applying scDAC to large-scale scRNA-seq datasets, researchers can consider increasing the number of network layers, the dimensions of each hidden layer and the dimensionality of the low-dimensional representation. Additionally, the number of training epochs and the minibatch size can also be appropriately increased. Conversely, when dealing with small input datasets, the opposite changes can be considered. Regarding the DPMM module, the number of components α is typically set to 1e-10, which is robust. Other parameters in the DPMM module follow the default values recommended in the BayesianGaussianMixture function.
3.Please train the model:
$ CUDA_VISIBLE_DEVICES=0 py run.py --task [datasetname]
The parameters in scDAC include the network parameters in the AE module and the model parameters in the DPMM module. The parameter details of scDAC are as follows: in the AE module, we set the sizes of two hidden layers of encoder to 256 and 128, set the sizes of the hidden layers of decoder to 128 and 256, and set the dimensionality of the low-dimensional representation to 32; in the DPMM module, the number of components α is robust and often set to 1e-10, the other parameters are set to the default value suggested in the BayesianGaussianMixture function; in the loss function, we set the weight hyperparameter λ to 1 in all experiments. During the training process, the minibatch size is set to 512, and the AdamW optimizer with a fixed learning rate of 1e − 4 is used for optimization.
4.To infer the labels after training, run the following command:
$ CUDA_VISIBLE_DEVICES=0 py run.py --action infer_latent --task [datasetname] --init-model sp_latest
Then we can obtain the predict label and ARI, NMI and SC sore.
5.The indicator of Deviation Ratio (DR) can be calculated by running scDAC/metrics_DR.ipynb