Use and understanding of electroporation have grown in recent years, revolutionizing various fields. However, optimization for stimulation techniques is still needed. In this context, the introduction of genetically modified cell cultures allows a dramatic increase in simulation capabilities, but also introduces the necessity of more advanced and human independent analysis methods.
We aimed to identify features, including morphological characteristics and other experimental parameters, to develop models for predicting the responses of genetically engineered HEK (Human Embryonic Kidney) cells to pulsed electric fields.
This subset of predictive features, including the presence of K+ channels, electric field strength, experiment number, initial fluorescence, and cell morphology characteristics was identified.
A machine learning approach based on ensemble learning techniques deployed through the XGBoost algorithm was utilized. This approach involves sequentially building numerous weak decision trees, where each subsequent tree aims to correct the errors made by the ones before it. Considering the unbalanced frequencies of the cell response types, we adopted different strategies to balance the training set and avoid bias were adopted.
The produced XGBoost model trained with a combination of real and synthetic data exhibited an accuracy of 66.0%, a mean AUC of 0.89, and an average F1 score of 0.66 when evaluated against the internal test set comprising solely real data. Further analysis on an external test set revealed an F1 score of 0.57.
In conclusion, we identified predictive features and produced models that may contribute to predicting the responses of genetically engineered HEK cells to pulsed electric fields.
