Emerging Safety Science: Workshop Summary (2008)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

9 Integration A s the workshop made clear, much of the work being done in safety science is focused in two areas: understanding and predicting toxicity in the discovery and preclinical stages of development, and spotting increased risk as soon as possible in the postmarket stage. As Janet Woodcock, Deputy Commissioner and Chief Medical Officer, U.S. Food and Drug Administration (FDA), noted in her introductory comments, however, the two areas should not be treated as separate spheres. Indeed, she said, one of the most significant challenges of emerg- ing safety science will be to bring the two together. Ultimately, safety sci- ence research must be iterative, with insights from one area being used to provide direction for investigations in the other. The long-term vision is to have all areas of safety science fully integrated, from the discovery through the postmarket stages. To that end, several speakers described various ways in which such integration is being accomplished now and offered visions of what it might look like in the future. An integration tool at GLAXOSMITHKLINE Almenoff noted that although the FDA receives about 400,000 adverse event reports each year for marketed products, there is very limited sys- tematic feedback of that clinical safety information to the discovery pipe- line. And while there is likely a great deal of information embedded in This section is based on the presentation of Dr. Almenoff. 93 

94 EMERGING SAFETY SCIENCE those reports that could be of value to researchers developing drugs, these data have not been leveraged. For example, adverse event reports may contain information that could help in discerning which chemical struc- tures may and may not be associated with particular problems. Almenoff described one approach that GlaxoSmithKline (GSK) has taken to attempt to extract and utilize that unrevealed information. Molecular Clinical Safety Program GSK developed the Molecular Clinical Safety Program (MCSP) as a way of closing the knowledge gap among various disciplines and help- ing to minimize both safety risks and attrition in the drug pipeline. The program consists of a large data warehouse that will ultimately house information on about 21,000 compounds, plus a set of tools for working with those data. Each compound is anchored to a chemical structure that can be linked to all the information in the database: class and dose infor- mation, physical properties, toxicology, pharmacokinetics, and metabo- lism and bioassay data. Each compound is also linked to human safety data, including disproportionality analysis (DPA) scores calculated from data in the FDA’s Adverse Event Reporting System (AERS), drug labeling information, and literature submissions. Ideally, the warehouse will also contain data and results from clinical trials, but because those data are more difficult to include, the system does not yet contain them. The tools for interrogating the data include various statistical tools, such as recursive partitioning, as well as molecular mapping, visualiza- tion, and query and search tools. Starting with a particular drug, for example, it is possible to obtain its structure and then ask the database for a list of all other compounds with a similar structure, or perhaps all other compounds containing a particular substructure that can be used to compare a drug of interest with a reference drug. After a list of com- pounds has been generated, bioassay data can be obtained on all of those compounds, or the evaluator can examine those compounds that are on the market and evaluate their safety data. It is also possible to query for all the compounds that bind a particular molecule and examine the safety data associated with those molecules. Example: Nelarabine and Neurotoxicity Almenoff offered an example of how GSK researchers used this query tool to answer a retrospective question about nelarabine (a chemotherapy drug marketed by GSK as Arranon) that had been shown to cause demy- elination in primate studies. Researchers wanted to determine whether interrogating the MCSP database could have provided them with infor- mation that might have made them proceed differently. 

INTEGRATION 95 As noted, the MCSP database allows researchers to enter a drug of interest and ask for a listing of all similar compounds and any relevant toxicities. In this case, the researchers asked the system for marketed drugs with a chemical structure similar to that of nelarabine and any toxicities associated with those drugs. For three of the six compounds that resulted from the similarity search, there were serious neurotoxic- ity signals—such as demyelination and polyneuropathy—rated on the EBGM (empirical Bayes geometric mean) scale Almenoff had described earlier (see Chapter 8). As it turned out, nelarabine did show some dose-related clinical neu- rotoxicity during development, but it is approved for use as a second- or third-line drug for refractory T cell lymphomas and leukemias. While it does have a very favorable risk/benefit balance for lymphoma and leukemia, it carries a black box warning for the demyelinating toxicity. If the MCSP tool had been available during the development of nelarabine, researchers might have seen the potential neurotoxicity problem early on and might have made a different decision about the drug’s development, perhaps deciding to proceed with a backup drug, for instance. Even hav- ing early knowledge of the neurotoxicity issues, researchers might have decided to move forward with the compound because they believed it was still the best option; even so, however, they would have been aware of the types of outcomes that would ultimately need to be monitored for. Almenoff argued that the MCSP is one more tool in the decision-making process: it does not necessarily cause a change in the development path of a drug, but it does provide more information for use in the decision- making process. Example: A Receptor Associated with Tardive Dyskinesia A great many preclinical screens exist, and more are continually being developed. An important question is which of these screens are most pre- dictive of clinical toxicity. Almenoff offered an example of how the MCSP tool was used to model toxicity data by linking human clinical safety data with molecular targets. GSK researchers examined tardive dyskinesia (TD), which is a major adverse effect of schizophrenia treatment. Working from the entire AERS database, they screened for bioassay results that were most strongly asso- ciated with TD. The analysis was performed with 600 marketed drugs because those were the drugs for which GSK had compound profiling data at the time. Working from the EBGM scores that measured the drugs’ statistically estimated risk factor for TD, the analysis looked for assays that discriminated between drugs with high and low TD signals. They found that certain catecholamine receptor subtypes, particularly at very high potency, were extremely strong predictors for TD. Thus, this analysis 

96 EMERGING SAFETY SCIENCE was able to pinpoint the relationship between catecholamine inhibition and TD. But the analysis also identified a strong association with another receptor. Although it is commonly believed that the action at the dopa- mine receptors explains the connection between antipsychotic drugs and TD, the analysis found six to eight compounds for which there was a high signal score, but the signal was better explained by this second receptor. Almenoff warned that while the findings are still preliminary, the researchers believe they have discovered an association between a recep- tor and TD that was previously unknown. Antipsychotic drugs bind with this second receptor very avidly; furthermore, this receptor is localized in an area of the brain that modulates movement. Although this associa- tion has not been verified, the important point is that the MCSP tool can be used to discover new and potentially valuable information with data that already exist. the elsevier database (PHARMAPENDIUM) Dr. MacLaughlin described a different approach to integration: an information tool that integrates preclinical, clinical, and postmarket safety data and makes it possible to look for patterns and connections among these data. In 1999 he was the principal investigator in a cooperative research and development agreement between Elsevier and the FDA to predict toxicological and adverse event end points. Leveraging historical data, the researchers sought to establish a comprehensive database com- prising well-organized and integrated preclinical, clinical, and postmarket data using FDA approval packages. Accumulating the Data FDA approval packages, or summary basis of approval packages, consist of the medical reviews, pharmacology reviews, and summaries of other data collected on a compound reviewed during the FDA’s approval process. These packages offer a rich source of information, but because of the format in which the FDA presents them publicly (e.g., nonindexed, nonsearchable paper; microfiche; bitmap formats), they are not readily accessible. Therefore the data cannot be queried for such information as class, target, and effect. To transform the data into a more usable format, Elsevier scanned roughly 750,000 pages character by character. Given the complexity of the data, it was necessary to have MDs and PhDs review This section is based on the presentation of Philip MacLaughlin, Senior Product Manager, Pharmaceutical Development, Elsevier. 

INTEGRATION 97 and interpret them once they had been scanned. While this was a difficult and labor-intensive process, the final product yielded data in a format that could be accessed as a modern electronic document (i.e., it could be copied, searched, and manipulated electronically). The resulting database included records dating back to 1992 and sometimes earlier, totaling about 35,000 approval packages. Organization and Context One barrier encountered by the researchers was inconsistencies in the terminology used throughout the approval packages. Different packages used different terms for the same thing. For example, “electrocardiogram QT prolonged,” “long QT,” “QT increased,” “QT interval prolonged,” “prolonged ventricular repolarization,” and “increased QT interval” are just a few of the terms used in the approval packages for a prolonged QT interval. Overcoming this barrier required a careful review of each pack- age, followed by mapping of each term to a standard term, such as that given in MedDRA, the Medical Dictionary for Regulatory Activities. Another barrier involved hierarchical terms. For example, if one is interested in retrieving data about ventricular arrhythmias from the data- base, it will be important for the database to recognize that there are many different types of ventricular arrhythmias: premature ventricular contrac- tion, multifocal ventricular tachycardia, wide complex ventricular tachy- cardia, ventricular flutter, etc. This barrier can be overcome by organizing all of the terms into hierarchical thesauruses, so that it is possible not only to know when two terms mean the same thing, but also to understand their relationships—when one is a restricted case of another, for example. Such thesauruses had to be defined for all the different types of terms found in the approval packages, including adverse events, drugs, and targets. Overcoming this terminology barrier was a difficult process, and creating these thesauruses required qualified and properly trained scien- tists, of whom there is currently a shortage. Once the thesauruses were completed, however, the team had a way to classify and find relationships for every term found in the approval packages (see Figure 9-1). The lack of standardized terminology throughout the drug development industry will continue to pose a major barrier to the integration of datasets. Strategic organization of the database is critical. One method for effi- ciently organizing the data is to use drug names and structural chemistry as the foundation for anchoring all other data. If this method is used in conjunction with the drug thesauruses described above, then regardless of what compound name is queried—generic or trade name—the correct compound with its structure and all the other links to related informa- tion will be retrieved. Once the thesauruses had been built and the data 

98 AE/Tox DRUG TARGET (MedDRA) (w/structural chemistry) FIGURE 9-1  Example of hierarchical thesauruses for adverse events, drugs, and targets. NOTE: MedDRA = Medical Dictionary for Regulatory Activities. SOURCE: MacLaughlin, 2007. Landscape View 

INTEGRATION 99 extracted from the now-digitized approval packages, the resulting data- base, with all the original information in standardized form, could be interrogated. MacLaughlin described the work of a pharmaceutical company that looked for comparative exposure data, particularly in humans, but also in animals. After 3 years of combing the literature and validating experi- ments in an effort to create its own database of a single end point, the company had identified a total of 600 drugs with this end point. If the database described above had performed this search, it would have been possible to retrieve within moments data on more than twice as many drugs—1,400 instead of just 600—and with a full population of param- eters, not just a single data point. Summary In summary, MacLaughlin offered several take-away messages and next steps: • There is no easy solution to the integration of data sets; however, properly planning for the future will facilitate the effort. The key is to have a uniform, standardized database spanning the entire development process. • The lack of standardized terminology throughout the drug devel- opment industry and the health care community is a major barrier. • Using this database, it is possible to look at preclinical, clinical, and postmarket data simultaneously; identify all compounds with a certain substructure and a certain target; and list all the toxicological effects of a particular type. • The database described above is available to reviewers at the FDA today, but more generally, its creation can serve as a template for similar efforts—for example, for the collection and organization of the kinds of toxicogenomics data described in Chapter 4. an fda perspective Dr. Frueh offered an FDA perspective on emerging safety science, with a particular focus on what is needed to integrate the various stages of drug development. Within the field, he explained, there are two main questions the FDA is interested in answering: At the preclinical stage, is it possible for tests to screen out compounds that have the potential This section is based on the presentation of Felix Frueh, Associate Director for Genomics, Office of Clinical Pharmacology, Center for Drug Evaluation and Research, FDA. 

100 EMERGING SAFETY SCIENCE to induce toxicity in humans? And at the clinical stage, is it possible to develop tests or diagnostics that can measure the probability of drug- induced toxicity? As described by the workshop’s presenters, researchers are working to improve safety at each stage of drug development. At the preclinical stage, for example, gene expression analyses are being used to predict toxicity. In addition to various technologies under development, Frueh emphasized that it will be vital to develop feedback loops through which information from one stage of development is used to make safety assess- ments at another. At each stage of development, researchers should be thinking about how to apply the information generated at that stage to early development so they can establish better tests, enhance the use- fulness of biomarkers employed in the drug development process, and improve decision making. The Need for Bridging Biomarkers While it is important to create feedback loops that enable data from later in the process to inform technologies and decisions in earlier stages, it is also important to create forward connections. Frueh stressed the importance of having bridged biomarkers from the preclinical to the clinical stage in the event that preclinical information could predict what might happen in the clinical stage (see Figure 9-2). While a number of traditional nonclinical test systems are used to assess safety—animal tests, tissue slices, cell cultures, in silico models, etc.—they are not always good predictors of toxicity in humans. Further- more, many of these systems rely on the signal generated by a toxicity state, and it would be preferable to have tests that can identify toxicity before it manifests and is detectable by histopathology assessments. In contrast to these traditional methods, the biomarkers described by many speakers during the workshop are, in a sense, surrogates for toxicity—they indicate when toxicity will develop but before it actually does so. A number of such markers are available today, and many others are being developed. While they may work well for compound selection and early characterization, however, they are not necessarily good predic- tors of a toxic event in humans. Creation of a more efficient drug development paradigm will require the establishment of biomarkers that can bridge or translate early preclini- cal findings to the clinical stage. These biomarkers would be of the same type in preclinical models and in humans and would represent quantifi- able indicators of normal biological processes, pathophysiological states, and responses to therapeutics. In some fields, studies to identify and vali- date such bridging biomarkers are already under way. For example, the 

INTEGRATION 101 Postmarket Development and use of a diagnostic test New biomarker discovery and optimization Prototype Clinical Development FDA Filing/ Basic Preclinical Design or Approval & Research Development Discovery Phase 1 Phase 2 Phase 3 Launch New biomarker discovery and optimization Bridging Markers Development and use Premarket Toxicogenomic and other “ -omic” of a diagnostic test profiling for prediction of toxicity Classification of compounds FIGURE 9-2  Drug development as an iterative process. The figure illustrates how biomarkers may be used to bridge, or translate, early preclinical findings to clini- cal findings, and how clinical findings may be used to inform and corroborate the basic science. Many of the workshop participants emphasized that the ultimate goal of applying these new technologies in9-2 science is to create a continual Figure safety iterative process in which the basic scientific data can help to inform and predict clinical outcomes. Revised SOURCE: Frueh, 2007. Predictive Safety Testing Consortium, a partnership between the FDA’s Critical Path Initiative and a number of large pharmaceutical companies that is aimed at predicting the safety of new treatments before their use in humans, is investigating the analytical validation of a set of nephrotoxic- ity markers, and in the near future it will consider methods for establish- ing clinical qualification. Such bridging biomarkers would likely be highly useful in explor- atory IND (investigational new drug) work. If these markers are indeed mechanistic and provide information about underlying toxicity from a molecular mechanistic point of view, it should be possible to discern tox- icity signals very early in humans—before harm to organs occurs—and at very low doses. Frueh hypothesized that the availability and measure- ment of a variety of bridging markers could well result in the creation of a 

102 EMERGING SAFETY SCIENCE completely new “phase 0” safety package. Although this idea is at present only conceptual, he believes there is true potential for the application of such bridging biomarkers in very early-stage drug development. Dealing with Idiosyncratic Events Frueh explained that in addition to developing biomarkers to predict preclinical and clinical events, it is necessary to establish biomarkers for dealing with idiosyncratic events—random, unexpected, often dose- independent adverse drug reactions—that occur during clinical testing and once a drug is on the market. These events are generally caused by an interplay between the properties of a drug and the predispositions— genetic and otherwise—of the patient. In such cases, the only option is to learn after the fact, studying the event to understand what caused it. Withdrawal of drugs from the market because of the occurrence of idiosyncratic events is harmful not only to patients suffering the adverse events, but also to patients not at risk who would otherwise benefit from a drug. However, the only way to reduce the occurrence of idiosyncratic events is to develop processes and invest in research that can lead to a reduction in all adverse events, both serious and nonserious. As an exam- ple of this approach, Frueh described work on drug-induced long QT syndrome. This is an idiosyncratic, rare event, but after hepatotoxicity, it is the top reason for drug withdrawals. The effect is generally reversible—if a patient stops taking the drug in question, he or she reverts to a normal QT—but it can sometimes lead to torsades de pointes, a condition that can be fatal. While not all QT prolongation leads to torsades de pointes, it is impossible to predict when this will occur, so it is necessary to regard QT prolongation as a marker for the potential development of that fatal effect. Since drug-induced long QT syndrome occurs in a wide variety of structurally diverse compounds, it is impossible to make a class pre- diction. Even though researchers believe they understand some of the mechanisms involved—namely, many of the drugs that induce long QT syndrome are KCNH2 (HERG) blockers—other factors clearly play a role as well. For instance, many drugs that block the same channel do not induce long QT syndrome, and therefore it is not easy to predict which drugs will do so. To predict which drugs may induce long QT, researchers would need to design a study that could identify new genetic biomarkers that could be used to determine whether a drug had the potential to cause prolonged QT. Such a study would need to consider the influence of external factors, such as medications or other exposures, but it would also need to look closely at genetic factors. Although some people have a genetic predispo- 

INTEGRATION 103 sition to congenital long QT syndrome, the same mutation can also lead to different phenotypes. Some with the mutation have QT prolongation in the absence of drugs, while others have it only in the presence of a drug; this genetic mutation is responsible for about 10 percent of cases of drug- induced long QT syndrome. The study would also need to consider other relevant genotypes, such as CYP2D6 and drug-metabolizing enzymes. The bottom line is that many genetic factors likely play a role, and it will be necessary to identify these factors and perhaps formulate some risk pattern that would make it possible to assess an individual’s risk of experiencing long QT syndrome when given various drugs. Frueh hypoth- esized that most likely, a genome-wide single nucleotide polymorphism (SNP) analysis conducted in a large number of patients using a variety of different drug classes and taking all other factors into consideration will be required to develop a hypothesis about the phenotype–genotype asso- ciation underlying this phenomenon. Once this has been accomplished, the association can be qualified with a separate data set, and ultimately, causation can be established, molecular mechanisms mapped out, and the resulting understanding applied in the clinical setting to avoid drug- induced long QT syndrome in as many patients as possible. Monitored Release Regardless of how proficient scientists become with bridging bio- markers and understanding of idiosyncratic events, it will never be possible to know for certain whether a drug is totally safe. The main problem is that the safety databases generated during drug develop- ment are generally too small to highlight rare events successfully. The largest amount of safety data is actually produced once a drug is on the market, but current tools do not allow scientists to capture this informa- tion effectively and capitalize on its potential. Typically, there are tens of patients in phase I trials, tens to hundreds in phase II, and hundreds to thousands in phase III. If an adverse event occurs in 1 of every 5,000 or 10,000 patients, it may be impossible to detect such an event prior to the drug’s approval. Frueh suggested that the problem could be addressed by instituting a system of monitored release (see Figure 9-3), which would be invoked after a drug’s initial but before its final approval. In such a system, the first 100,000 patients (or whatever number was selected) to take a new drug would be monitored for adverse events. Samples would be collected from any patients who experienced such events, along with samples from a matched group of controls, and these samples would be analyzed to identify the factors leading to the adverse event. Once it was possible 

104 EMERGING SAFETY SCIENCE Monitor the first e.g., 100,000 patients that receive the drug, collect samples from patients experiencing an AE and from matched controls, conduct e.g., WGA to identify genetic basis Safety Biomarker Characterization for AE and what could be done to prevent it in future Monitored Full Exploratory (Learn) Validation (Confirmatory) Release Release Initial Full Approval Approval Modeling and Simulation Continuous Interaction with health authorities FIGURE 9-3  Illustration of a product development timeline that includes a moni- tored released phase. This phase would be invoked after a drug’s initial but before Portrait view its final approval. Throughout monitored release, samples would be collected from patients to enable study of the genetic basis of adverse events. Note: AE = adverse event; WGA = whole genome association [studies]. SOURCE: Frueh, 2007. Figure 9 -3 to identify at-risk patients and remove them from the population being prescribed that drug, the drug could proceed to final approval. Summary Frueh summarized the potential of these emerging technologies and some of the future barriers scientists must overcome: • The emergence of new molecular biomarkers for drug safety will make it possible to better bridge the safety gap between the preclinical and clinical stages; the hope is that eventually, having true translational biomarkers will transform the process into a continuum. • Because a drug’s safety can never be completely proven, research- ers will always have to rely on the absence of signals; however, the pro- cess of looking for such signals can be significantly improved. • The development of better characterizations for toxicity will lead 

INTEGRATION 105 to better markers for toxicity, but accomplishing this will require a true interdisciplinary approach involving experts in all relevant fields. • Emerging safety science has already made it possible to better clas- sify compounds through new genomic and other technologies. • Researchers need to qualify markers for bridging studies, as well as those for addressing idiosyncratic events. • Nephrotoxicty biomarkers will be submitted to the FDA for review this year, and a process for reviewing these markers is being established. • Genomic association studies (including, for example, whole-genome SNP scanning) have the potential to identify markers for rare adverse events, but access to well-characterized samples remains a problem. • New mechanisms and processes for studying clinical and postmar- ket safety need to be explored. Following the workshop, the Predictive Safety Testing Consortium’s nephrotoxicity bio- marker package was submitted to the FDA for review.