OR WAIT null SECS
A new paradigm that could transform a drug failure into a drug success would narrow the target population to those patients most likely to benefit from such treatment.
Each year, many promising therapeutic agents intended to address some of our most serious unmet medical needs fail in clinical development due to the lack of efficacy or unacceptable toxicity. Inevitably, many of these agents offer significant clinical benefit for some patients. However, they often fail to demonstrate sufficient safety and efficacy across a larger patient population to the degree needed to surmount the high clinical and economic thresholds often established by pharmaceutical companies, regulatory agencies, and the marketplace.
Figure 1. Evolutionary Drug Developmentâ¢, a proprietary iterative process to identify optimal rules.
There is a growing body of evidence and increasing acceptance that many of these promising compounds can be made commercially and economically viable. By identifying patient subpopulations that are most likely to benefit from and/or least likely to adversely react to these therapeutic agents, researchers could "personalize medicine." This targeted approach holds the promise of more rapidly and safely bringing both old and new medicines to those patients most in need.
Among the medicines that have become successfully commercialized consequent to identifying a responsive patient subpopulation is trastuzamab (Herceptin), an anticancer monoclonal antibody that targets the HER-2 growth factor receptor on breast cancer cells. It is approved as a single agent for second-line treatment and in combination with paclitaxel for first-line treatment of patients with metastatic breast cancer whose tumors specifically overexpress the HER-2 protein.1 Had this responsive subpopulation not been prospectively identified during the clinical development of trastuzamab, the drug would likely not have been approved in breast cancer, consequently depriving many women of a treatment that has proven to have prolonged many of their lives.
On the other side of the equation are drugs that are efficacious in many patients but are not tolerated by some. Thalidomide (Thalomid) is the classic example of a medicine that has been commercially resurrected subsequent to its complete withdrawal from the world market (in 1961) for unacceptable toxicities. Initially promoted as a sedative and treatment of such conditions as asthma, hypertension, and migraine, it soon came into disfavor consequent to an associated epidemic of severe and often fatal perinatal malformations, most frequently phocomelia. Capitalizing on its diverse beneficial effects as an immune modulator and the successful identification of those patients for which it is contraindicated for safety reasons, thalidomide is now successfully overcoming its stigma as an unacceptably toxic compound and is showing promising activity in a wide variety of solid malignant tumors, multiple myeloma, graft-versus-host disease, and Crohn's disease.2
Clozapine is another example of a medicine that has found a therapeutic niche despite being associated with significant toxicity, in this case life-threatening agranulocytosis. Clozapine was approved in 1989 for second-line symptomatic management of schizophrenia and is now being investigated for use in Parkinson's disease, Huntington's chorea, and depression.3
The key to the successes of thalidomide and clozapine has been the ability to prevent their distribution to patients identified to be most at risk for their toxic side effects, namely pregnant women and those with severe granulocytopenia, respectively. In both cases, mandatory controlled distribution programs are in effect, helping to ensure the safe and effective prescribing, dispensing, and use of these two medicines.3,4
There are three key challenges to introducing a personalized medicine approach to clinical practice: identifying the optimal patient subpopulation for receipt of a specific therapeutic agent, based on potential benefit-versus-risk; developing rapid, economical, accurate, and clinically practical methods for determining whether or not a specific individual belongs to such a patient subpopulation; and ensuring distribution of the agent to those for whom the agent will be potentially most effective and nontoxic.
Although all disciplines of medicine could potentially benefit from the personalized medicine approach, this paper will concentrate on the challenges being faced in clinical oncology—a field that is fostering and benefiting from many emerging technologies and in which molecularly targeted therapies are showing great promise.
What can be done to ensure that there is a high likelihood that the patients entered in a clinical trial will respond to and tolerate a particular anticancer agent? Traditionally, this has been addressed by incorporating fairly standard patient inclusion and exclusion entry criteria in Phase II and III clinical trial designs. To some extent, this narrows the patient population down to including those that have at least some chance of responding while excluding those that are likely not to tolerate the treatment. Although this routinely generates an enrolled study population that would be commercially attractive if enough patients responded, the response rate is usually not high enough to convince most drug developers and regulatory agencies that marketing it for that indication is warranted.
This is a major cause for the high rate of perceived "drug failures" in oncology drug development and a source of much frustration for both investigators, patients, and the general public. A new clinical trial design paradigm that could potentially transform a drug failure into a drug success would be one that continues to narrow the target population to those patients most likely to benefit from such treatment while also continuing to tolerate it.
Many novel molecularly targeted therapies are being developed in oncology. Although it is assumed requisite for anti-tumor activity that a compound's cellular molecular target be present on or in the patient's tumor, the degree to which this target needs to be expressed remains unclear. Although overexpression of HER-2 is required for trastuzamab to be clinically beneficial in metastatic breast cancer, the same appears not to be the case for those agents targeting the epidermal growth factor receptor (EGFR), such as gefitinib (Iressa) and erlotinib (Tarceva). For gefitinib in advanced non-small cell lung cancer (NSCLC), the presence of genomic mutations in the EGFR appears to correlate with gefitinib-induced tumor regression.5 However, in a large Phase III placebo-controlled registrational trial of gefitinib, such biological responses were relatively uncommon and did not translate into an overall statistically significant survival benefit for the advanced NSCLC patient population studied.6 In contrast, erlotinib has been shown to produce a significant prolongation of survival in advanced NSCLC.7
In similar NSCLC patient populations, why did erlotinib demonstrate prolonged survival while gefitinib did not? The answer is likely hidden within the complexity and heterogeneity of tumor cells and the diverse biologies, demographics, and medical histories of cancer patients—all interacting to determine an individual's ultimate response to a particular anticancer agent.
To answer such a question for oncology compounds in general, we are testing the hypothesis that by integrating the vast array of existent baseline patient and tumor information that is usually readily retrievable from concluded Phase II and/or Phase III studies of that agent, one can identify particularly sensitive, insensitive, tolerant, and/or intolerant patient subpopulations. Examples of such baseline patient and tumor data currently available include: patient demographics, medical history, response to prior cancer therapy, as well as tumor type, stage, and histology. Baseline tumor and patient genomic and protein biomarker data will become increasingly available in the near future and hold promise to significantly enhance the accuracy of such subpopulation assessments.
To proceed with such investigations, the key question of what are the most clinically meaningful outcomes for the oncology compound under study first needs to be answered. Phase II trial designs may range from single-arm, uncontrolled studies that depend on comparison of results with historical controls to multi-arm, controlled studies that have a standard-of-care arm for comparison, as with Phase III studies. Depending on the intent of the trial, the stage and natural history of the cancer, and the mechanism of action of the study drug, clinically meaningful efficacy endpoints might include survival (e.g., one-year, overall, etc.), time-to-disease progression, the incidence of stable disease (as defined in the protocol), quality-of-life, and/or changes in cancer-related biomarkers, e.g., serum PSA in prostate cancer.
The goal is to generate "outcome predictors" (a.k.a. "screening classifiers"), that can be statistically validated in subsequent proof-of-concept Phase II clinical studies. If shown to have adequate positive and negative predictive values, these new "rules" for patient selection can then be incorporated into larger targeted Phase III proof-of-efficacy trials. Because trials designed to incorporate such outcome predictors will require fewer randomized patients than untargeted designs, such an approach could potentially conserve significant amounts of drug development resources.8
A major challenge to generating useful outcome predictors in this fashion, both now and in the future, is how to efficiently condense and analyze the ever increasing amount of baseline data obtained from a relatively small number of patients. Such information is derived from diverse technologies like quantitative immunohistochemistry, flow cytometry, polymerase chain reaction, fluorescent in situ hybridization, and gene expression profiling.
Whereas traditional statistical analysis, cluster analysis, and neural networks are available to tackle such analyses, their utility will become more limited as the number and diversity of input variables increase. An analysis technique that is particularly capable of handling such voluminous data is a machine learning approach called "genetic programming."9 Borrowing ideas from natural selection and population dynamics, genetic programming is an iterative analytical process that searches for computer programs that are good enough to survive in competition with others in a population of programs. The result is a set of computer programs that solves a particular problem, in this case the prediction of clinical response and/or tolerance to a specific therapy.
Evolutionary Drug Development™ is a process developed by Genetics Squared that uses a proprietary implementation of genetic programming to describe a set of data, gaining insight into the data and suggesting how a drug can be best used for a particular patient population.10 This iterative process of analysis, research, and refinement affords powerful advantages over other approaches, namely: 1) the ability to combine very large and diverse data types (e.g., genomic, proteomic, medical history, etc.); 2) the ability to find novel combinations of factors; 3) the ability to combine variables in nonlinear ways; and 4) the generation of simple and readable rules that may support known mechanisms of action for specific targeted compounds and/or suggest novel molecular interactions that may warrant further investigation (Figure 1).
A recent Phase II study conducted by a major pharmaceutical company had failed to meet internal thresholds for success with approximately 25 percent of patients experiencing stable disease (SD), as defined in the protocol and appropriately reflecting the natural history of the cancer under study. Team statisticians at the company had been unable to correlate the occurrence of SD with any of the biomarkers or other clinical variables captured.
The pharmaceutical company wanted the data analyzed to see if there was a correlation. They supplied to us approximately 30 demographic and medical history variables, efficacy outcomes, and four types of tumor biomarker values obtained by immunohistochemistry. Employing Evolutionary Drug Development™, a predictive rule was discovered very early in the analysis of baseline information that was greater than 70 percent accurate in predicting whether or not a patient would achieve SD (unpublished results). This rule used a single tumor biomarker that the clinical team had assumed was not important because it was highly expressed in all patients' tumors. Our rule discovered a difference between high and very high expression of this biomarker. While the rule was very good at predicting who would not respond (i.e., specificity), it was poor at predicting positive responses (i.e., sensitivity).
If the drug were being developed to treat a non-life threatening disease and was known to have safety issues, then this would have been an excellent rule. However, because the compound was tolerable and was being developed for metastatic cancer, this rule would miss too many patients that could benefit from it. Therefore, subsequent data analyses were focused on identifying a new rule that optimized sensitivity versus specificity.
The new predictive rule combined medical history and demographic variables with two biomarkers to produce a much more compelling rule with similar overall accuracy, but sensitivity over 90 percent. If this rule had been used to screen patients for the clinical trial, it would have missed only two patients and resulted in a 60 percent incidence of SD in this patient subpopulation—a compelling outcome for a compound being developed for a metastatic solid tumor. The pharmaceutical company was excited enough by these results to request reviews of two additional Phase II studies of this compound in other tumor types, as well as a clinical study of another oncology compound. These studies are ongoing.
Challenges inherent in the practical implementation of personalized medicines include:
1) producing and validating reliable pharmacodiagnostic tools that can be used by community physicians to help them formulate treatment strategies for individual patients; and 2) convincing regulatory authorities that this novel method of patient selection will ensure adequate levels of patient safety. Matching patients to the most appropriate and promising therapeutic agents must be rapid, economical, accurate, and clinically practical.
Aside from reviewing responses to prior treatment modalities, patients and their tumors are tested against a variety of predictive rules and panels of biomarkers. These rules and biomarker data might be housed in large registries that can be accessed in real-time by the prescribing physician or dispensing pharmacist via the Internet or telephone to maximize distribution of drug to "safe responders."
The clozapine patient registry managed by Outcome houses predictive rules along with specific patient information that permits ongoing checks to be made for specific patterns of response or toxicity each time a prescription is refilled. Using registries such as this as a model will help to ensure the continued safe and appropriate use of drugs after approval. Such successful integration of predictive rules and biomarkers with information technology permits the reliable point-of-care access that is key to enabling personalized medicine to quickly move from the laboratory to the clinic.
The era of personalized medicine is here. Many medicines that have had their clinical development discontinued or that have been withdrawn from the marketplace due to lack of perceived benefit or excessive toxicity may be able to be rescued and successfully developed if a sensitive and tolerant patient subpopulation can be identified. By finding new and better defined niches for medicines of all types, many novel compounds will become available to those patients that are most likely to benefit from them, while also permitting the use of quicker and less expensive clinical trials during their development.
Evolutionary Drug Development™ is an implementation of genetic programming that is being used at the forefront of this effort, one well positioned for meeting the challenges of making sense of the exponentially increasing amount of available patient and tumor data that needs to be analyzed.
Finding the right rules and markers is but the first critical step. The next step, moving personalized medicine from the laboratory to the clinic, will require increasingly sophisticated methods to deliver the right information rapidly and reliably to prescribers and pharmacists. To this end, real-time accessible, computational patient registries that are focused on appropriate drug distribution are already in use. These, together with cutting-edge technological approaches like Evolutionary Drug Development™, are spearheading the transition of personalized medicine from bench to bedside and hold much promise for rescuing many valuable compounds along the way.
1. Herceptin (Genentech): Full Prescribing Information, revised October 2003.
2. M.E. Franks, G.R. Macpherson, W.D. Figg, "Thalidomide," Lancet, 363 (9423)1802-1811 (2004).
3. Clozapine (Mylan): Full Prescribing Information, revised January 2003.
4. Thalomid (Celgene): Full Prescribing Information, revised February 2004.
5. T.J. Lynch et al., "Activating Mutations in the Epidermal Growth Factor Receptor Underlying Responsiveness of Non-small-cell Lung Cancer to Gefitinib," New England Journal of Medicine, 350 (21) 2129-2139 (2004).
6. Doctor Letter Regarding Iressa (gefitinib) ISEL Data (December 17, 2004): A double blind, placebo controlled, parallel group, mulitcentre, randomised, Phase III survival study comparing IRESSA (gefitinib) (250mg tablet) plus best supportive care versus placebo plus best supportive care in patients with advanced NSCLC who have received one or two prior chemotherapy regimens and are refractory or intolerant to their most recent regimen. AstraZeneca Pharmaceuticals LP, 2004.
7. Tarceva (Genentech; OSI Pharmaceuticals): Full Prescribing Information, initial November 2004.
8. R. Simon and A. Maitournam, "Evaluating the Efficiency of Targeted Designs for Randomized Clinical Trials," Clin Cancer Res, 10, 6759-6763 (2004).
9. J.R. Koza, Genetic Programming (MIT Press, Cambridge, MA, 1992).
10. D. MacLean, E.A. Wollesen, W.P. Worzel, "Listening to Data: Tuning a Genetic Programming System," Genetic Programming Theory and Practice II, U.-M. O'Reilly, T. Yu, R. Riolo, W.P. Worzel, eds. (New York, Springer Science+Business Media, Inc., 2005) pp. 245-262.
Peter Lenehan, MD, PhD, is chief medical officer, Bill Worzel is chief technology officer, and John Freshley is president and CEO of Genetics Squared, 210 South Fifth Avenue, Suite A, Ann Arbor, MI 48104, (734) 929-9475, fax (734) 929-9477, email; email@example.com. Richard Gliklich, MD, is president of Outcome, 210 Broadway, Cambridge, MA 02139, (617) 621-1600, fax (617) 621-1620.