Classification of unknown primary tumors with a data-driven method based on a large microarray reference database

Abstract

We present a new method to analyze cancer of unknown primary origin (CUP) samples. Our method achieves good results with classification accuracy (88% leave-one-out cross validation for primary tumors from 56 categories, 78% for CUP samples), and can also be used to study CUP samples on a gene-by-gene basis. It is not tied to any a priori defined gene set as many previous methods, and is adaptable to emerging new information.

Figure 1

A graph of the accuracy of the method as a function of the similarity score of the best hit. The graph was formed by moving a sliding window of width 0.1 along the score axis, which ranges from -0.021 to 0.495, and calculating the achieved accuracy within that window. As can clearly be seen, the better the similarity, the higher the probability that the classification is correct.

Figure 2

Similarities for 83 metastatic test samples. (a) A comparison of the test samples' similarities to the healthy tissue where the metastasis was found (y-axis) and the cancer of origin for the metastasis (x-axis). The spheres are colored according the site of the metastasis ('target tissue'). The gray diagonal line indicates a boundary, above which the similarity to the target tissue is greater than the similarity to the sample's original cancer. Only ten samples display this behavior, and all but one of these are lymph node metastases. (b) A comparison of the test samples' similarities to a representative cancer of the tissue where the metastasis was found (y-axis) and the cancer of origin for the metastasis (x-axis). The triangles are colored according to the site of the metastasis ('target tissue'). The gray diagonal line indicates a boundary above which the similarity to the cancer of the target tissue is greater than the similarity to the sample's original cancer. As can be seen, most samples fall below this line.

Figure 3

Cancer-specific genes. (a) Tissue specificity scores, unmodified by gene weight, for genes whose weight in renal cancer is greater than 0.25 (40 out of 58 present) are shown for 10 renal cancer metastasis samples. The genes can be divided into two groups, those that lose their renal cancer-specific expression (blue) and those that do not (red). The samples are named according to where the metastasis was located, and numbered according to their (relative to each other) similarity to renal cancer. Sample 10 was the only one whose closest similarity was not renal cancer, it instead being lung squamous cell carcinoma. Samples 1 to 3 are the closest to renal cancer, and retain for most of the genes renal cancer-specific expression levels. The other samples have lost renal cancer-specific expression among the genes with a blue background. (b) Similar analysis for the 17 metastatic melanoma samples, showing 42 (out of 63) genes. (c) Similar analysis for 10 metastatic gastric cancer samples, showing 40 (out of 53) genes.