To train and evaluate TACTiCS, use the TACTiCS class in tactics.py. As input, TACTiCS requires two scRNA-seq datasets and a distance matrix describing the gene similarities. calc_dist calculates the latter for embeddings created with embed_proteins in genes.py. The preprocessed data, trained model and resulting cell embeddings can be saved using TACTiCS.save.
For full description, please check the function descriptions.
TACTiCS requires PyTorch. The transformers package is only necessary for embedding protein sequences.
The tutorial.ipynb notebook explains step-by-step how to use TACTiCS for a comparison across two species, including how to create protein sequence embeddings.
All datasets used are publicly available. For convenience, the protein sequences, protein embeddings, raw count data and trained models for 50k subsets of the data can be found on Zenodo. The full unprocessed dataset can be downloaded here.