This is the official implementation of the paper : "Visualizing hierarchies in scRNA-seq data using a density tree-biased autoencoder"
All requirements are found in the file requirements.txt.
You will additionaly have to install Speedrun. An installation guide is available at github.com/inferno-pytorch/speedrun.
The training takes place in two parts, first the pretraining and then the finetuning.
We assume that the data is named X.csv and its folder is defined in the experiment's configuration.
In order to train a model, you can either create manually a configuration folder/files like in the folder experiments and then use the commands:
python <pretrain/finetune>.py <experiments_folder>/<experiment_name>
Or you can use an already existing experiment and use :
python <pretrain/finetune>.py <experiments_folder>/<experiment_name> \
--inherit <experiments_folder>/<reference_experiment_name>
Notice that here that the folder <experiment_name> doesn't need to exist already
To modify this existing experiment you can use the command line parameters
--config.<config_element> <new_element_value>
As such, modifying the dimensions of the autoencoder in an experiment would look like
python <pretrain/finetune>.py <experiments_folder>/<experiment_name> \
--inherit <experiments_folder>/<reference_experiment_name> \
--config.model.kwargs.dims [<dim_1>,<dim_2>,...,<dim_n>]
For more details regarding Speedrun's usage, please refer to github.com/inferno-pytorch/speedrun.
After the training, the final embedding is in the file experiments/experiment/embedding.csv. Embeddings are also stored periodically during training, and logs are visualizable during training using Tensorboard. They are found in the folder experiments/experiment/Logs.
We provide both template experiments and the configurations used to reproduce our results in folders experiments/<experiment>-<pretraining/finetuning>.
Pretrained networks used to produce the figures in the paper are available on Google Drive
The preprocessed endocrine pancreas dataset is available at : GSE 132188
The raw dentage gyrus dataset is available at : GSE 104323
We will take a look at how to reproduce the experiment on the PHATE generated data.
The process is as follows :
python phate_data_generation.py
python pretrain.py experiments/phate-gen-pretraining --config.device cuda:0
python finetune.py experiments/phate-gen-finetuning --config.device cuda:0
python embedding_visualization.py --path experiments/phate-gen-finetuning
The first line creates the dataset and stores it in the folder data.
The second one will pretrain the DTAE on the cgenerated data(data/X.csv), on device cuda:0. Feel free to use the appropriate device here.
Then the third will finetune the DTAE and produce the final embedding.
The fourth line is here to visualize in a very simple way the embedding obtained during training.