We introduce scMultiNODE, an unsupervised integration model that combines gene expression and chromatin accessibility measurements in developing single cells, while preserving cell type variations and cellular dynamics. First, scMultiNODE uses a scalable, Quantized Gromov-Wasserstein optimal transport to align a large number of cells across different measurements. Next, it utilizes neural ordinary differential equations to explicitly model cell development with a regularization term to learn a dynamic latent space. (bioRxiv preprint)
If you have questions or find any problems with our codes, feel free to submit issues or send emails to jiaqi_zhang2@brown.edu or other corresponding authors.
Our codes have been tested in Python 3.7. Required packages are listed in ./installation.
-
Raw and preprocessed data of six temporally resolved multi-modal single-cell datasets can be downloaded from here (for HC, HO, DR, and MN data) and here (for ZB and AM data).
-
All model integrations and corresponding evaluation metrics on six datasets are available at here (modal_integration.zip).
-
Investigation of scMultiNODE with cell type supervision are available at here (cell_type_supervision.zip).
-
Experiment results for downstream analysis (cell trajectory pseudotime estimation, cell path construction, and cross-modal cell label transfer) are available at here (downstream_analysis.zip).
-
Data visualization and figures in the paper are available at here (journal_figs.zip).
-
scMultiNODE is implemented in ./model/dynamic_model.py.
-
All baseline codes are provided in baseline. See the documentation therein for more details.
The script of using scMultiNODE for integration is shown in ./modal_integration/Modal_Integration_scMultiNODE.py.
- data: Scripts for data preprocessing. Some scripts are implemented in R and need installation of Seurat.
- model: Implementation of scMultiNODE model.
- optim: Loss computations, QGW algorithm, evaluation metrics.
- baseline: Implementation of baseline models.
- modal_integration: Run each model on six multi-modal single-cell datasets and compute integrations. Evaluation metrics computation and comparison.
- downstream_analysis: Use scMultiNODE for cell trajectory pseudotime estimation, cell path construction, and cross-modal label transfer.
- hyperparameter_investigation: Ablation study and investigation of hyperparameter settings for scMultiNODE.
- tuning: Hyperparameter tuning for scMultiNODE and baselines.
- plotting: Visualization / figure plotting.
- utils: Utility functions.
Please report any bugs, problems, suggestions, or requests as a Github issue or send emails to jiaqi_zhang2@brown.edu or other corresponding authors.