senkin-tmp-cite-pred

Reimplementation of the model for surface protein prediction from transcriptomics data, which won OpenProblems 2022 competition in CITE-seq task.

Installation

Create a conda environment and install the package:

conda create -n cite_pred python=3.13 -y
conda activate cite_pred
pip install git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@main

Faithfulness to the original solution

Version 0.2.0 fixes several deviations from the original pipeline (see CHANGELOG.md). The preprocessing was validated against the feature matrices of the original solution computed from the competition raw counts: the CLR-TSVD components, the custom normalization with its TSVD/PCA components, and the selected correlated genes are reproduced. Remember that the neural networks expect every feature block to be z-scored per cell, which senkin_tmp_cite_pred.nn_models.prepare_nn_inputs does for you.

Running the preprocessing and the model

See example.ipynb for a detailed description and an example on how to run the model.

See open_problems_example.ipynb to see how to run the model with OpenProblems data format.

• Github repository: https://github.com/lueckenlab/senkin-tmp-cite-pred/