- Selects a **minimal set of genes** that can discriminate two different status in the selected cell population(s)

- Selects a **minimal set of genes** that can discriminate two different status in the selected cell population(s)

- Trains **patient-level ML/DL classifiers** that can predict patients in two different status

- Trains **patient-level ML/DL classifiers** that can predict patients in two different status

Specifically, patients are split into training and testing sets. In the training set, cell populations responsive to perturbations are scored by quantifying how well each cell population is separated between two conditions. With the selected population, Support Vector Machine Recursive Feature Elimination (*SVM*-*RFE*) is applied to identify a minimal number of genes with high predictive power. The number of genes in the panel is automatically decided in a data-driven way to avoid bias from manual inspection. Using the selected cell population(s) and corresponding gene panel(s), scPanel constructs a patient-level classifier with the training data and evaluates its performance in the testing data to validate the power of identified genes. All the data splitting involved in scPanel is done at the patient level so that the importance of the selected cell population, genes, and the performance of corresponding classifiers are genearalizable to all patients.

Specifically, patients are split into training and testing sets. In the training set, cell populations responsive to perturbations are scored by quantifying how well each cell population is separated between two conditions. With the selected population, Support Vector Machine Recursive Feature Elimination (*SVM*-*RFE*) is applied to identify a minimal number of genes with high predictive power. The number of genes in the panel is automatically decided in a data-driven way to avoid bias from manual inspection. Using the selected cell population(s) and corresponding gene panel(s), scPanel constructs a patient-level classifier with the training data and evaluates its performance in the testing data to validate the power of identified genes. All the data splitting involved in scPanel is done at the patient level so that the importance of the selected cell population, genes, and the performance of corresponding classifiers are genearalizable to all patients.

### 💡 Why scPanel is better:

### 💡 Why scPanel is better: