Figure 3

QDA derived outlier scores in CUP. A) To determine the relationship between prediction error and outlier scores the primary cancers and metastases were divided into ten bins according to the outlier scores and the error rate was calculated for each bin. Each point represents the error rate plotted versus the median outlier score of the bin. The vertical lines show the span of outlier scores within the bins. The plot shows that higher outlier score translates into higher error rate. We modeled the relationship between outlier scores and prediction error by fitting polynomial function to the data points (the orange line), and the function allows us to estimate the expected error rate for new samples of unknown origin, once their outlier scores have been determined. B) Samples from CUP patients tend to have higher outlier scores than other cancer patients. The box plot summarizes the distributions of outlier scores within metastases (MET), primary (PRIM) and CUP tumors. There is a clear tendency for CUP samples to have higher outlier score than metastases and primary cancers. The median outlier score of CUP samples of >1000 suggests the origin prediction error above 30%. On the other hand, most primary cancers and metastases have outlier scores below 800, hence the estimated prediction error from 2-10% (see panel A). Since data for CUP and some primary tumors and metastases were generated at Rigshospitalet, the non-CUP samples from Rigshospitalet are presented as separate group (RH_MET and RH-PRIM), this is to show that the shift in outlier scores was not caused by technical bias. Additionally, the normal, non-cancerous tissue group (NORMAL) is included, and shows the whole range of outlier scores.