A working conda environment could be downloaded. Please follow the README.txt file there to activate this environment. Once successfully activated, you may skip the installation steps.
2022/05/03 New features experimented, details introduced in
notebooks
2022/04/29 VeloAE updated to 0.2.0. To simplify the project, the folder
notebooks
is reorganized and only notebooks involving veloAE experiments are kept, scvelo dynamical mode models are additionally included for comparison. The previous data are backuped in the branchpaper-version-backup
2022/04/29 Version Updating in progress. We thank @Mingze Yuan from PKU for his great insights in correcting issues regarding veloAE's cohort aggregation module and a suggestion on replacing GCN with GAT layers, which leads to better performances on challenging datasets like human and mouse bonemarrow, a preview of updated results:
2022/04/04 Exciting news! UnitTVelo, a new single cell RNA velocity estimation tool that addresses the challenging datasets of existing tools is published by our lab.
Low-dimensional Estimation of Single Cell RNA Velocity with AutoEncoder.
VeloAE can learn low-dimensional projections for count matrices leveraging a tailored AutoEncoder, with the aim to obtain better representations for RNA velocity estimation. Results of VeloAE could be previewed in the jupyter notebooks located under the notebooks
folder.
Our revised manuscript is in progress, while the first version could be found in bioarxiv with the title Representation learning of RNA velocity reveals robust cell transitions
Follow https://github.com/qiaochen/VeloAE/blob/main/notebooks/readme.md to access the datasets used in the notebooks and manuscript
Install
pip install git+https://github.com/qiaochen/VeloAE
conda create -n veloAE
conda activate veloAE
git clone https://github.com/qiaochen/VeloAE.git
cd VeloAE
conda install python=3.7
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip install torch-sparse -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip install torch-geometric
pip install .
Basic execution command and arguments:
veloproj --refit false --vis_type_col clusters --scv_n_jobs 10 --vis-key X_umap --nb_g_src X --gumbsoft_tau 5 --fit_offset_pred true --adata notebooks/dentategyrus/data/DentateGyrus/10X43_1.h5ad --device cuda:3 --model-name ./notebooks/dentategyrus/dentategyrus_model.cpt
Arguments:
Output:
veloproj --lr 1e-5 --nb_g_src X --gumbsoft_tau 5 --fit_offset_pred true --vis_type_col clusters --scv_n_jobs 10 --vis-key X_umap --refit true --adata notebooks/dentategyrus/data/DentateGyrus/10X43_1.h5ad --device cuda:3 --model-name dentategyrus_model.cpt --output './'
Arguments:
Output:
training_loss.png
in the output folder (default ./, can be specified using arg --output).If the training loss decreases smoothly but not converge after exhausting all the epochs, please try either more epochs or larger learning rates. Note, however, that learning rates should not be set too large to make the loss osciliates and never decreases.
Use command line help to investigate more arguments.
veloproj -h