Questions about drug perturbation prediction using cell line data

[GA-1101]

Hi,
I used MIX-Seq results data for drug perturbation of different cell lines to train the SCGEN model and select the indicated cell lines for prediction. After testing, the following problems were found:
The correlation between the predicted result obtained by the prediction and the control result is high, while the correlation between the predicted and the stimulated result is low, and the correlation between the control and stimulated is also low. We speculate that SCGEN may have failed to predict drug perturbations in specific cell lines, thus maintaining a high degree of correlation in very unmedicated samples (ie, control).

The data we used is from MIX-Seq.

The correlation results is following:
predict vs stimulated:

predict vs control:

control vs stimulated: