Fig. 1: Evaluating different methods for classifying single-nuclei transcriptomes by disease status.
From: Neural networks reveal novel gene signatures in Parkinson disease from single-nuclei transcriptomes
A Schematic of the analysis workflow. For this analysis, we used publicly available single-nuclei RNA sequencing (snRNAseq) data from the post-mortem substantia nigra of individuals with Parkinson disease and controls prepared by Kamath et al. Four feature selection methods were independently applied to the snRNAseq data before being input into four distinct machine learning (ML) models for disease classification of individual cells. A total of 16 unique combinations of feature selection methods and ML models were tested, with each combination applied to classify seven distinct cell types independently and all cell types together. For each feature selection-ML model combination, we trained five independent models per cell type using five-fold cross validation (n = 128 models). B Boxplot showing the disease classification accuracy obtained using different feature selection methods. C Boxplot showing the disease classification accuracy obtained using different ML models. A Wilcoxon rank-sum test was used to compare mean balanced accuracies; non-significant comparisons are not shown. NN neural network, HVG highly variable gene, LR logistic regression, PC principal component, RF random forest, SVM support vector machine. *** P < 0.001; **** P < 0.0001.
