Concise prediction models of anticancer efficacy of 8 drugs using expression data from 12 selected genes.

We developed concise, accurate prediction models of the in vitro activity for 8 anticancer drugs (5-FU, CDDP, MMC, DOX, CPT-11, SN-38, TXL and TXT), along with individual clinical responses to 5-FU using expression data of 12 genes. We first performed cDNA microarray analysis and MTT assay of 19 human cancer cell lines to sort out genes which were correlative in expression levels with cytotoxicities of the 8 drugs; we selected 13 genes with proven functional significance to drug sensitivity from a huge number of potent prediction marker genes. The correlation significance of each was confirmed using expression data quantified by real-time RT-PCR, and finally 12 genes (ABCB1, ABCG2, CYP2C8, CYP3A4, DPYD, GSTP1, MGMT, NQO1, POR, TOP2A, TUBB and TYMS) were selected as more reliable predictors of drug response. Using multiple regression analysis, we fixed 8 prediction formulae which embraced the variable expressions of the 12 genes and arranged them in order, to predict the efficacy of the drugs by referring to the value of Akaike's information criterion for each sample. These formulae appeared to accurately predict the in vitro efficacy of the drugs. For the first clinical application model, we fixed prediction formulae for individual clinical response to 5-FU in the same way using 41 clinical samples obtained from 30 gastric cancer patients and found to be of predictive value in terms of survival, time to treatment failure and tumor growth. None of the 12 selected genes alone could predict such clinical responses. Copyright 2004 Wiley-Liss, Inc.