sICTA 0.0.1

A marker-based cell type annotation method that combines the self-training strategy with pseudo-labeling and the nonlinear association capturing capability of Transformer.

sICTA: Interpretable Cell Type Annotation based on self-training

The full description of sICTA and its application on published single cell RNA-seq datasets are available.

Download archive with preprocessed data at: https://drive.google.com/drive/folders/1jbqSxacL_IDIZ4uPjq220C9Kv024m9eL.

The repository includes detailed installation instructions and requirements, scripts and demos.

1 The workflow of sICTA.

(a) Combining cell expression and marker gene specificity to generate pseudo-labels. (b) The downstream Transformer classifiers are first pre-trained based on cell type probability distributions (pseudo-labels), followed by iterative refinement of the classifiers through a self-training framework until convergence. The sICTA takes the a priori knowledge from the biological domain and uses masked learnable embeddings to transform the input data ($G$ genes) into $k$ input tokens representing each gene set (GS) and a class token (CLS).

2 Requirements

• Linux/UNIX/Windows system
• Python == 3.8.6
• torch == 1.12.1
• scanpy == 1.9.1

3 Usage

Data format

sICTA requires cell-gene matrix and cell type marker information to be entered in .h5ad object format.

Training

python main.py

We provide default data for users to understand and debug sICTA code.

Reference

If you use sICTA in your work, please cite