TOSICA: Transformer for One-Stop Interpretable Cell-type Annotation

Package: TOSICA

We created the python package called TOSICA that uses scanpy ans torch to explainablely annotate cell type on single-cell RNA-seq data.

Requirements

• Linux or UNIX system
• Python >= 3.8
• torch >= 1.7.1

Create environment

conda create -n TOSICA python=3.8 scanpy
conda activate TOSICA
conda install pytorch=1.7.1 torchvision=0.8.2 torchaudio=0.7.2 cudatoolkit=10.1 -c pytorch

Installation

The TOSICA python package is in the folder TOSICA. You can simply install it from the root of this repository using

pip install .

Alternatively, you can also install the package directly from GitHub via

pip install git+https://github.com/JackieHanLab/TOSICA.git

Environment has been tested

TOSICA.yaml

Usage

Step 1: Training the model

TOSICA.train(ref_adata, gmt_path,label_name=<label_key>)

Input:

• ref_adata: an AnnData object of reference dataset.
• gmt_path : default pre-prepared mask or path to .gmt files.
• <label_key>: the name of the label column in ref_adata.obs.

Pre-prepared mask:

• human_gobp : GO_bp.gmt
• human_immune : immune.gmt
• human_reactome : reactome.gmt
• human_tf : TF.gmt
• mouse_gobp : m_GO_bp.gmt
• mouse_reactome : m_reactome.gmt
• mouse_tf : m_TF.gmt

Output:

• ./mask.npy : Mask matrix
• ./pathway.csv : Gene set list
• ./label_dictionary.csv : Label list
• ./weights20220603/ : Weights

Step 2: Prediect by the model

new_adata = TOSICA.pre(query_adata, model_weight_path = <path to optional weight>)

Input:

• query_adata: an AnnData object of query dataset .
• model_weight_path: the weights generated during scTrans.train, like: './weights20220607/model-6.pth'.

Output:

• new_adata.X : Attention matrix
• new_adata.obs['Prediction'] : Predicted labels
• new_adata.obs['Probability'] : Probability of the prediction
• new_adata.var['pathway_index'] : Gene set of each colume
• ./gene2token_weights.csv : The weights matrix of genes to tokens

Warning: the var_names (genes) of the ref_adata and query_adata must be consistent and in the same order.

query_adata = query_adata[:,ref_adata.var_names]

Please run the code to make sure they are the same.

Example Demo:

Guided Tutorial