Fig. 4

The curves of Pearson’s correlation coefficients and losses in the training of the HASCAD models. Each model was trained by using a set of 8000 simulated RNA-seq samples, and was validated by using another set of 8000 simulated RNA-seq samples. (A) and (B): The training data were simulated from the 3’v1 (threepv1) and 3’v2 (threepfresh) sets. The validation data were simulated from the 5’(fivePrime) set. (C) and (D): The training data were simulated from the 3’v1 (threepv1) and 5’ (fivePrime) sets. The validation data were simulated from the 3’v2 (threepfresh) set. (E) and (F): The training data were simulated from the 3’v2 (threepfresh) and 5’ (fivePrime) sets. The validation data were simulated from the 3’v1 (threepv1) set. In (A), (C), and (E), the scRNA-seq data were normalized by harmony. In (B), (D), and (F), the scRNA-seq data were not normalized by Harmony. The training was stopped when the model was not improved in loss function for the evaluation after 20 epochs