Figure 6.
Assigning single-cell gene set activity scores using an autoencoder. The autoencoder is designed such that the decoder is composed of binary vectors with values indicating if a profiled gene belongs to a known gene set or not. The output of the encoder, r, serves as coefficients for the gene set vectors, showing how related a cell is to a known pathway/biological process. r thus can be seen as the gene set activity scores for this cell. The set cover loss is designed to select uncorrelated pathways/processes to better annotate cells. (FC layers) Fully connected layers, (Bio-process) biological process.
