Tutorials

We provide an example going through how to use scSemiProfiler to preprocess and semi-profile a small dataset with 12 COVID-19 samples from patients of 6 different severity levels (stored in the example_data folder in our GitHub repository). To semi-profile a cohort, the following steps will be executed: (1) initial setup, which includes preprocessing and clustering bulk data, and selecting initial representatives; (1.5) obtaining single-cell data for representatives; (2) processing single-cell data and performing feature augmentations; (3) single-cell inference using deep generative models.

Then, once the inference is complete, the semi-profiled cohort can be utilized for various single-cell-level downstream analyses and compared with the results of the real-profiled cohort. The high similarity between the real and semi-profiled versions demonstrates the reliable performance of scSemiProfiler. If the budget allows, you have the option to employ an active learning algorithm to select additional representatives and proceed to the next round of semi-profiling. As more representatives are selected, the semi-profiling performance typically improves, but the costs also increase. We illustrate this trade-off relationship with an overall error versus cost curve.

You can also download our GitHub repository and run the example locally. Before running the notebook, your need to install scSemiProfiler and then install the conda environment as a Jupyter Notebook kernel:

conda install ipykernel
python -m ipykernel install --user --name=semiprofiler --display-name="scSemiProfiler"

Then open the notebook. You can now select the kernel “scSemiProfiler” in Jupyter Notebook and run our example notebook. Instructions of running Jupyter Notebook can be found here.

• Example
  • Step 1 Initial Setup
  • Step 1.5 Acquiring Single-cell Data for Representatives
  • Step 2 Single-cell Processing & Feature Augmentation
  • Step 3 Single-cell Inference
    • 3.1 Single target sample inference
    • 3.2 Inference sanity check
    • 3.3 Inference for the entire cohort
  • Comprehensive evaluation using downstream tasks
    • Assemble semi-profiled cohort
    • Read the real-profiled single-cell data to compare
    • Compare the UMAP visualization
    • Compare cell type composition
    • Compare gene set activation pattern
    • Compare top cell type signature genes
    • Use RRHO plot to compare markers
    • Compare GO enrichment analysis similarity
    • Compare partition-based graph abstraction (PAGA) graph similarity
    • Using CellChat to perform cell-cell interaction analysis
    • Export single-cell data for analysis
  • End of comprehensive semi-profiling performance evaluation
  • Optional: New Representative Selection and Run the Next Round
  • Round 2 semi-profiling
    • select a new batch of representatives using active learning:
    • obtain single-cell data for new representatives:
    • run single-cell inference again with more representatives:
  • stop criteria
  • new representative selection using active learning
  • Round 3 semi-profiling
  • Round 4 semi-profiling
  • Error curve
    • compute errors of each round
    • visualize the errors