TY - JOUR
T1 - Modeling tissue contamination to improve molecular identification of the primary tumor site of metastases
AU - Vincent, Martin
AU - Perell, Katharina
AU - Nielsen, Finn Cilius
AU - Daugaard, Gedske
AU - Hansen, Niels Richard
PY - 2014/5/15
Y1 - 2014/5/15
N2 - Motivation: Contamination of a cancer tissue by the surrounding benign (non-cancerous) tissue is a concern for molecular cancer diagnostics. This is because an observed molecular signature will be distorted by the surrounding benign tissue, possibly leading to an incorrect diagnosis. One example is molecular identification of the primary tumor site of metastases because biopsies of metastases typically contain a significant amount of benign tissue. Results: A model of tissue contamination is presented. This contamination model works independently of the training of a molecular predictor, and it can be combined with any predictor model. The usability of the model is illustrated on primary tumor site identification of liver biopsies, specifically, on a human dataset consisting of microRNA expression measurements of primary tumor samples, benign liver samples and liver metastases. For a predictor trained on primary tumor and benign liver samples, the contamination model decreased the test error on biopsies from liver metastases from 77 to 45%. A further reduction to 34% was obtained by including biopsies in the training data.
AB - Motivation: Contamination of a cancer tissue by the surrounding benign (non-cancerous) tissue is a concern for molecular cancer diagnostics. This is because an observed molecular signature will be distorted by the surrounding benign tissue, possibly leading to an incorrect diagnosis. One example is molecular identification of the primary tumor site of metastases because biopsies of metastases typically contain a significant amount of benign tissue. Results: A model of tissue contamination is presented. This contamination model works independently of the training of a molecular predictor, and it can be combined with any predictor model. The usability of the model is illustrated on primary tumor site identification of liver biopsies, specifically, on a human dataset consisting of microRNA expression measurements of primary tumor samples, benign liver samples and liver metastases. For a predictor trained on primary tumor and benign liver samples, the contamination model decreased the test error on biopsies from liver metastases from 77 to 45%. A further reduction to 34% was obtained by including biopsies in the training data.
U2 - 10.1093/bioinformatics/btu044
DO - 10.1093/bioinformatics/btu044
M3 - Journal article
C2 - 24463184
SN - 1367-4803
VL - 30
SP - 1417
EP - 1423
JO - Bioinformatics
JF - Bioinformatics
IS - 10
ER -
