We used Single Cell Clustering Assessment Framework (SCCAF)27 to test the robustness of our MSN subtype classification. The concept behind such 'self projection' tests is that the gene expression patterns in a subsample of the cells should be sufficient to classify the remaining cells in the labeled clusters with a high level of accuracy.85 SCCAF splits the expression data into a training and test set, and then fits a classifier on the training set on each cluster provided (MSN subtypes in our case). We used the default parameters of a 50% train/test split and a logistic regression classifier. We also used SCCAF to test whether our snRNA-seq sampling was sufficient for accurately detecting the MSN subtype heterogeneity. Because the number of cells of each MSN subtype could vary during the down sampling, we opted to use SCCAF classifier's macro-f1 score for evaluation. The f1 score is the geometric mean between the precision and recall. To compute our macro-f1 score, we computed the average across f1 scores for each MSN subtype. Using the scanpy.pp.subsample method, we repeatedly subsampled fractions of our cells from both monkeys and calculated the macro-f1 across each trial.
