In this module, we extract patterns to discern between cells with WT and Null genotypes using cell morphologies from plates 3, 3 prime, and 5. After substantial development, we have developed and optimized our final NF1 model with the same architecture as a logistic regression. The training and testing sets were chosen to create an approximately 90-10 percent train-test split between wells, while retaining a uniform genotype class distribution of cells within each data split. This minimizes the loss of cells when downsampling to the minority genotype class in each data split. We optimize the chosen model by performing a random search with stratified k-fold cross validation using the class-balanced training set. After optimization, we retrain the NF1 model with the optimal regularization strength using the entire dataset, but excluding the test set.
During optimization of the NF1 model, we save the predicted genotype probabilities, true genotype labels, and the predicted genotype labels of the training, validation, and testing datasets for later evaluation.