Conventional development of multivariate gene expression models (GEM) predicting therapeutic response of
cancer patients is based on analysis of patients treated with specific regimens, which limits generalization to different or novel
drug combinations. We overcome this limitation by developing GEMs based on in vitro
drug sensitivities and microarray analyses of the NCI-60
cancer cell line panel. These GEMs were evaluated in blind fashion as predictors of
tumor response and/or patient survival in seven independent cohorts of patients with breast (n = 275), bladder (n = 59), and ovarian (n= 143)
cancer treated with multiagent
chemotherapy, of which 233 patients were from prospectively enrolled clinical trials. In all studies, GEMs effectively stratified
tumor response and patient survival independent of established clinical and pathologic
tumor variables. In
bladder cancer patients treated with neoadjuvant
methotrexate,
vinblastine,
Adriamycin (
doxorubicin), and
cisplatin, the 3-year overall survival for those with favorable GEM scores was 81% versus 33% for those with less favorable scores (P = 0.002). GEMs for
breast cancer patients treated with
5-fluorouracil,
Adriamycin (
doxorubicin), and
cyclophosphamide and
ovarian cancer patients treated with
platinum-containing regimens also stratified patient survival [5-year overall survival 100% versus 74% (P = 0.05) and 3-year overall survival 68% versus 43% (P = 0.008), respectively]. Importantly, clinical prediction using our in vitro GEM was superior to that of conventionally derived GEMs. We show a facile yet effective approach to GEM derivation that identifies patients most likely to benefit from selected multiagent
therapy. Use of such in vitro-based GEMs may provide a robust and generalizable approach to personalized
cancer therapy.