Fig. 3
From: Identifying tumor cells at the single-cell level using machine learning
ikarus performance under perturbation. A Performance of ikarus classifier, on the Lambrechts’ validation set, when 20 to 80% of the input list is removed. ikarus performance is significantly reduced when the 60% of the gene list is removed (p value < 0.01, as determined by a one sample t-test, with BH-FDR correction). B Performance of ikarus classifier, on the Lambrechts’ validation set, upon single gene removal from the gene list. One hundred sixty out of the 162 genes were inconsequential for the classification. FGB (fibrinogen) and SAA1 (serum amyloid alpha) ablation significantly reduced the classification accuracy, but only on the Lambrechts et al. lung cancer dataset. C, D FGB and SAA1 are respectively strong markers for one of the cell state clusters in the lung cancer dataset