Fig. 2.

Clustering performance in the 4-left-out experiment. Each network is trained by leaving 4 cell types out (LPGO technique). The cell types which are left-out are randomly selected, and the procedure is repeated 20 times. After the neural network training is completed, the encoding (learned representation) is computed for the test (left-out cells) and used as input to the K-Means algorithm. The output is then evaluated using a comprehensive set of metrics (see “Materials and methods” section)