Skip to content
Neural networks for classification of single cell transcriptomics data
Jupyter Notebook Python Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
assets release Jul 29, 2019
configs release Jul 29, 2019
notebooks release Jul 29, 2019

scNym -- Classify Single Cells

scNym is a neural network model for predicting cell types from single cell profiling data (e.g. scRNA-seq) and deriving cell type representations from these models. While cell type classification is the main use case, these models can map single cell profiles to arbitrary output classes (e.g. experimental conditions).

We developed scNym as part of the Murine Aging Cell Atlas. If you find scNym useful in your work, please cite our preprint.

A murine aging cell atlas reveals cell identity and tissue-specific trajectories of aging
Jacob C. Kimmel, Lolita Penland, Nimrod D. Rubinstein, David G. Hendrickson, David R. Kelley, Adam Z. Rosenthal
bioRxiv 657726; doi:

If you have any questions or suggestions, please feel free to email me:

Jacob C. Kimmel


The scNym model is a simple neural network leveraging modern best practices in architecture design. The model has a user configurable number of hidden layers, each with a user configurable number of parameters. Raw inputs are first processed with a dropout layer to simulate technical noise commonly observed in single cell profiling experiments. Subsequent hidden layers leverage the Residual Block design popularized in Residual Networks and are paired with ReLU activation, Batch Normalization, and Dropout layers.


First, clone the repository:

$ export SCNYM_PATH=~/ # change this to any path you prefer
$ git clone $SCNYM_PATH

We recommend creating a virtual environment for use with scNym. This is easily accomplished using virtualenv or conda.

$ virtualenv scnym_env
$ source scnym_env/bin/activate
$ pip install -r $SCNYM_PATH/requirements.txt


$ conda create -n scnym_env
$ conda activate scnym_env
$ pip install -r $SCNYM_PATH/requirements.txt

Once the environment is set up, simply run:

$ python build install

After installation completes, you should be able to run scNym as a command line tool:

$ scnym --help



Models can be trained using the included command line interface, scnym.
The CLI accepts configuration files in YAML or JSON formats, with parameters carrying the same names as command line arguments.

To see a list of command line arguments/configuration parameters, run:

$ scnym -h

A sample configuration is included as default_config.txt.

Data Preprocessing

It is recommended that raw scRNA-seq counts be normalized prior to model training. A common approach normalized by the library size to Counts Per Million (CPM), then natural log + 1 transformed (i.e. x_norm = ln(x_cpm + 1)).

Input data can be stored as a dense [Cells, Genes] CSV of normalized counts, or in an AnnData h5ad object, or a Loompy loom object.

All of the below are valid:

scnym -c configs/default_config.txt --input_counts counts.csv ...
scnym -c configs/default_config.txt --input_counts counts.h5ad ...
scnym -c configs/default_config.txt --input_counts counts.loom ...

scNym models may either be trained on all genes, or a subset of pre-selected genes. To subset genes from the provided inputs matrix, simply use the --genes_to_use argument pointed at a newline delimited text file of gene names. The supplied --genes_to_use must be a subset of --input_genes.


scNym employs a notion of an "upper" and "lower" grouping for data. The upper grouping acts as additional information for model construction, while the lower groups corerspond to the output classes.

Commonly, the upper grouping may correspond to the tissue cells were isolated from, while the lower grouping corresponds to the cell types of interest.

scNym models are therefore available in three flavors: (1) tissue independent models, (2) tissue dependent models (conditional), and (3) tissue specific models.

Demo Script

A demo shell script is provided that downloads data from the Tabula Muris and trains an scnym model.

To execute the script, run:

chmod +x

in the repository directory.

The script is an example of how to train a model and can serve as a template for your own data.

Tissue Independent Model

Tissue independent models are trained on all cells provided without regard for the tissue of origin.

$ scnym train_tissue_independent -c configs/default_config.txt

Tissue Dependent Model

Tissue dependent models are conditioned on the tissue of origin for each cell profile.

This is implemented by appending a one-hot encoded categorical vector representing the tissue of origin to the single cell profile vector of each cell. This categorical vector is omitted from the initial dropout layer of the model.

$ scnym train_tissue_dependent -c configs/default_config.txt

Tissue Specific Training

Tissue specific models are trained on only one upper group (e.g. one tissue) at a time.

$ scnym train_tissue_specific -c configs/default_config.txt

Training Scheme

Class balancing is performed using both over- and undersampling. Minority classes with fewer than 128 examples were oversampled, while classes with >128 samples were undersampled.

Models are trained with 5-fold cross-validation using both a validation set for model selection and a hold-out test set for evaluation. Final models were trained on all available class-balanced data.

Output Description

Training a model with scnym with create the following directory structure:

    all_data/ # models trained on all data except for a validation set used for model selection
        # this directory contains the same structure as `fold00` below
        00_best_model_weights.pkl # the best model weights from this output fold
        labels.csv # labels for each sample
        predictions.csv # prediction for each sample when it was held-out
        test_idx.csv # samples for model selection
        train_idx.csv # samples for training
        val_idx.csv # held-out samples
        model_weights_NNN.pkl # model weights are saved at a specified frequency
        {EXP_NAME}_log.csv # a running log of losses and accuracies during training
        {EXP_NAME}_parameters.json # parameters provided to the model, JSON-ized
    celltype_label.csv # string labels for each output node
    fold_eval_acc.csv # accuracy for each fold of training
    fold_eval_losses.csv # loss for each fold of training
    # contains the same structure as `tissue_independent/`, but for 
    # the tissue conditional models
    # this directory is populated when using the `find_cell_type_markers` command
    {CLASS_INDEX}_{CLASS_NAME}_losses.csv # losses for input optimization of this class
    {CLASS_INDEX}_{CLASS_NAME}_optima.csv # optima for input optimization of this class
        # outputs for a classifier trained only on "lower groups"
        # in "UPPER_GROUP_NAME_0" are stored here.
        # contains the same structure as `tissue_independent/`

Trained Models

scNym models have been trained on (1) all cell types in the Tabula Muris data set without tissue information, (2) all cell types conditioned on the tissue of origin, and (3) separate models have been trained to predict cell types within each tissue.

We have provided pre-trained weights for models trained on the 10X data using the configuration in default_config.txt. Models were trained on all genes detected in an inner join of cells from the Tabula Muris.

To perform predictions with these models on new data, you'll need the list of gene names we used for training, as well as the training metadata (to get cell type output labels). We provide both below.

Training Data Gene Names

Training Metadata

Pre-trained Model Weights

10X Tissue Independent Model

10X Tissue Dependent Model

10X Tissue Specific Models

Pre-trained Model Performance

Reported accuracies are the mean accuracy across 5 cross-validation splits (seperate sets were used for model selection and evaluation).

Model Accuracy
Tissue Independent 0.957
Tissue Dependent 0.958
Bladder 0.992
Heart_and_Aorta 0.993
Kidney 0.983
Limb_Muscle 0.995
Liver 0.996
Lung 0.966
Mammary_Gland 0.985
Marrow 0.952
Spleen 0.956
Thymus 0.937
Tongue 0.964
Trachea 0.961

See the tutorial notebook for a demonstration of performance on a lung data set generated using microwell-seq technology, instead of 10X.


Once a model has been trained (or if you use one of the pre-trained models above), cell type predictions can be made using the CLI or an interactive interface, suitable for use in Jupyter notebooks.

Notebook Tutorial

In this tutorial linked below, we use a model trained on the Tabula Muris to predict cell types in the "Mouse Cell Atlas".

Notebook cell type prediction tutorial

CLI Prediction

CLI prediction uses the same configuration file interface as model training.

The two key parameters that differ are --training_gene_names and --model_path.

--training_gene_names is an ordered list of the gene names used in the original model training. Combined with the gene names you provide with the data set to be analyzed, scnym will sort or subset the genes to allow for prediction using the pre-trained weights. We've found that prediction still works well, even if only a subset of the genes used during training are present in the prediction data set.

--model_path is simply a path to the pickled model weights (as provided above).

scnym predict_cell_types \
    -c /path/to/config_file.txt \
    --training_gene_names /path/to/gene_names_used_for_training.txt \
    --model_path /path/to/model_weights.pkl
You can’t perform that action at this time.