PyTorch implementation of MARS, a meta-learning approach for cell type discovery in heterogenous single-cell data. MARS annotates known and new cell types by transferring latent cell representations across multiple datasets. It is able to discover cell types that have never been seen before and characterize experiments that are yet unannotated. For a detailed description of the algorithm, please see our manuscript Discovering Novel Cell Types across Heterogeneous Single-cell Experiments (2020).
MARS requires anndata and scanpy libraries. Please check the requirements.txt file for more details on required Python packages. You can create new environment and install all required packages with:
pip install -r requirements.txt
We implemented MARS model in a self-contained class. To make an instance and train MARS:
mars = MARS(n_clusters, params, labeled_exp, unlabeled_exp, pretrain_data)
adata, landmarks, scores = mars.train(evaluation_mode=True)
MARS provides annotations for the unlabeled experimentm as well as embeddings for annotated and unannotated experiments, and stores them in anndata object. In the evaluation_mode annotations for unlabeled experiment need to be provided, and they are used to compute metrics and evaluate the performance of the model.
MARS embeddings can be visualized in the 2D space using UMAP or tSNE. Example embeddings for diaphragm and liver tissues:
MARS can generate interpretable names for discovered clusters by calling:
mars.name_cell_types(adata, landmarks, cell_type_name_map)
Example of the MARS naming approach:
Example of running MARS on Tabula Muris dataset in leave-one-tissue-out manner is provided in the main_TM.py. We also provide two example notebooks that illustrate MARS on small-scale datasets:
- cellbench.ipynb demonstrates MARS on two CellBench dataset of five sorted lung cancer cell lines sequenced with 10Xand CEL-Seq2 protocols.
- kolod_pollen_bench.ipynb demonstrates MARS on Pollen dataset of diverse human cell types, and Kolodziejczyk dataset of mouse pluripotent cells.
We provide cross-validation benchmark cross_tissue_generator.py for classifying cell types of Tabula Muris data. The iterator goes over cross-organ train/test splits with an auto-download of Tabula Muris data.
Tabula Muris Senis datasets is from https://figshare.com/projects/Tabula_Muris_Senis/64982.
Tabula Muris Senis dataset in h5ad format can be downladed at http://snap.stanford.edu/mars/data/tms-facs-mars.tar.gz. Small-scale example datasets CellBench and Kolodziejczyk/Pollen can be downloaded at http://snap.stanford.edu/mars/data/cellbench_kolod_pollen.tgz.
Pretrained models for each tissue in Tabula Muris can be downladed from http://snap.stanford.edu/mars/data/TM_trained_models.tar.gz.
If you find our research useful, please consider citing:
@article{brbic2020mars,
title={MARS: Discovering Novel Cell Types across Heterogeneous Single-cell Experiments},
author={Brbic, Maria and Zitnik, Marinka and Wang, Sheng and Pisco, Angela O and Altman,
Russ B and Darmanis, Spyros and Leskovec, Jure},
journal={Nature Methods},
year={2020},
}
Please contact Maria Brbic at [email protected] for questions.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent