Response Profiling
Genomic Prognostic Models
Over the past decade, cancer investigators have identified a variety of genetic markers that may separate high-risk patients—those most likely to have early recurrence and cancer death after resection—from those who will not. Although aggressive tumors do not necessarily equate with tumors that respond to a particular chemotherapy regimen, in fact these high-risk patients have had the greatest benefit from adjuvant chemotherapy in most solid tumor systems. Initial work in this area has focused on refining predictive models of cancer recurrence after resection for patients with early-stage NSCLC.
Recently, the development of gene expression microarrays has provided us with the ability to record a "snapshot" of the active and inactive genes of a tumor. This tool has significantly changed the field of molecular prognostics and may allow future individual risk assessment. The major challenge is recognizing and defining useful data from "noise" as we sort through the microarray's thousands of simultaneous measurements. A new field of statistics, computation genomics, is developing novel methods to assist in this analysis.
For clinical application, a gene microarray-based prognostic model should be validated in a prospective, blinded manner. Recently, 4 studies have been published with some form of validation based on gene expression data.12,79-81 Unfortunately, the quality of the validation cohorts varied and only one investigator performed this validation in a completely blinded fashion.80 To build this type of a predictive model, a collection of gene expression profiles, termed metagenes, is generated. These computer-generated, randomly assigned sets of 25 to 200 genes are tested for their ability to separate patients based on survival. Once the computer identifies several hundred meta-genes that have prognostic significance, it begins to link them to select those that are additive in predictive value using a regression tree analysis. The ultimate regression tree in this study contained 100 metagenes, which combined to include more than 2000 distinct genes. This large size allowed for a predictive power much greater than previous investigations, which only included 30 to 50 genes.13,82-85 Ultimately, this model was validated on 3 external gene expression data sets of more than 120 patients, including blinded analysis of 2 National Cancer Institute Cooperative Group banked samples (American College of Surgeons Oncology Group and CALGB). This novel predictor is currently being used in an adjuvant chemotherapy clinical trial entitled CALGB 30506 "A Randomized Phase III Trial to Evaluate the Lung Metagene Score to Direct Adjuvant Therapy in Stage 1 NSCLC Patients."
Genomic Models for Selecting Therapies
Cytotoxic Chemotherapy
In lung cancer, response to cytotoxic chemotherapy has not improved in more than 10 years despite the introduction of several new agents. The core therapy is based on cisplatin, the most active single agent, paired with another drug. No combination has proven to be significantly different in terms of efficacy, so oncologists select therapy based on the toxicity profile of the combination. Genomic models are now being assessed for their ability to select the most appropriate chemotherapy regimen for a given patient, like a "bacterial antibiogram" used for infections. Two phase 2 feasibility studies are being activated that will select therapy based on a tumor profile: the Moffitt Cancer Center trial of gene expression of DNA repair gene ERCC1 (platinum resistance) and ribonucleotide gene RMM1 (gemcitabine resistance)86,87 and the Duke University trial using a gene expression profile developed from a lung cancer cell line to determine a tumor's platinum, taxane, and gemcitabine resistance.88 If successful, these will be validated in multi-institutional phase 3 trials (Figure 7).
Molecular Targeted Therapy
Today there are hundreds of selective molecular targets being developed and evaluated for US Food and Drug Administration approval and eventual use in the clinic. These agents have the potential to significantly change both oncology clinical practice and patient prognosis. Unfortunately, testing predictive models is difficult, both because an individual molecular target only has activity against specific components of an oncogenic pathway, which are usually seen in only a small fraction of patients with a particular tumor, and because no rational method yet exists to identify the "enriched" subset of patients whose tumors are most likely to have the specific oncogenic pathway inhibited by the target. One approach to address this challenge is a novel technique using gene expression data. It matches the gene signature of each patient's tumor with a profile of gene expression identified for an aberrant oncogenic pathway. These pathway profiles include 100 to 400 genes and represent components of all events in an entire pathway, not just an individual activating mutation.89 It is hoped that these pathway signatures will overcome the challenges in evaluating the efficacy of selective molecular targets in lung cancer.
Conclusion
Recent advances in genomic modeling and profiling have provided important new insights into cancer pathways and targets and are providing important new insights into risk assessment and drug response. Once validated in clinical trials, these genomic-based techniques will undoubtedly contribute significantly to individualizing and improving both assessment and outcomes for lung cancer patients.
Dr. Lynch: What do you think the differences will be between the Duke metagene profile and the 5-gene predictor from Taiwan?
Dr. Harpole: The value of the larger genomic gene sets is that they give you more discriminatory power. Some of the original ones had 20 to 30, and ours has 2100 genes. The bigger ones are going to have more power. Today it's easy to tell the difference between people with really bad cancers and people with really good cancers. Unfortunately, it's that 75% in the middle for whom these small predictors are useless.
Dr. Lynch: Do you see response profiling becoming a widely used tool for individualizing cancer therapy in the next decade?
Dr. Harpole: It will probably become much more widely used in cancer drug discovery and research, but routine clinical use to individualize therapy will probably require more time and testing.