Neuroscience Trials of the Future: Proceedings of a Workshop (2016)

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.

3 Clinical Trial Design Highlights • The National Institute of Mental Health (NIMH) established the Research Domain Criteria (RDoC) initiative to create a neurobi- ologically based research framework for addressing heteroge- neity (which has been shown to confound signal detection in Phase II trials) across and within disease entities (Morris). • There is a need to go beyond psychometric approaches and to- ward more causal modeling for nervous system disorders, using existing and new tools, in order to have more accurate endpoints in clinical trials (Bilder). • Candidate biomarker approaches have worked well in identify- ing biomarkers for Alzheimer’s disease that were later included in clinical trial enrollment criteria, but not as well for Parkin- son’s disease, where unbiased screening approaches have been more successful (Chen-Plotkin). • Biorepositories with strict collection and storage protocols, data sharing, and partnerships are essential for promising biomarkers to reach clinical trials (Chen-Plotkin). • Clinical as well as statistically significant evidence is important to establish safety and effectiveness for any given product; regu- lators and payers also look at clinical outcomes that are meaning- ful to patients (Bilder, Farchione, Hernandez, S. Kapur, Laughren, Pande, Pani, Peña, Pencina, Rockhold, and Romano). • Trial designs should be fit-for-purpose, and employ appropriate statistical expertise in the planning stage (Pencina). NOTE: These points were made by the individual speakers identified above; they are not intended to reflect a consensus among workshop participants. 15 

16 NEUROSCIENCE TRIALS OF THE FUTURE Clinical trials are typically conducted in three phases prior to market approval: Phase I, to evaluate safety, determine a safe dose range, and identify side effects; Phase II, to further evaluate safety and assess effi- cacy; and Phase III, to confirm effectiveness, monitor side effects, and compare it to other treatments (U.S. National Library of Medicine, 2008). The goal of Phase II is “to obtain preliminary data on whether the drug works in people who have a certain disease or condition.”1 Failures in Phase II are also of substantial concern, said Steven Romano; approx- imately 80 percent of the compounds tested in Phase II fail. More than half of these fail for lack of efficacy (Arrowsmith and Miller, 2013), and another 30 percent because of a company’s shift in strategic focus, which could reflect a company’s concerns about comparative effectiveness and thus marketability. The fundamental challenge, according to Robert Bilder, professor-in-residence in the department of psychiatry and biobe- havioral sciences at the University of California, Los Angeles, is the lack of validity for the “disease entities” associated with nervous system dis- orders, for which etiology and pathophysiology are mostly unknown. Both psychometric approaches and categorical approaches are insuffi- cient for most disease entities, and heterogeneity within a syndrome and across syndromes is common, which could confound signal detection in Phase II trials. RDoC, described below, is one attempt to address these problems. In addition, several participants noted the need for better dis- ease modeling—using both existing and new tools—and novel trial de- signs and statistical approaches. ALTERNATIVES TO THE DSM The DSM is a diagnostic tool that defines disorders as distinct enti- ties and relies mainly on subjective measures—self-report from patients and clinicians’ observations—rather than on an advanced understanding of the neurobiology of disease (Casey et al., 2013), said Steven Romano. In addition, the DSM fails to capture a number of disease domains such as volition, motivation, and speech (Kendler, 2016). Bilder added that categorical systems for defining disease are generally considered less valid than systems that capture the dimensionality of most psychiatric conditions (e.g., Haslam et al., 2012). He added that a dimension such as cognitive impairment may arise from different fundamental causes in 1 See http://www.adaptimmune.com/patients-families/clinical-trial-faq (accessed June 23, 2016). 

CLINICAL TRIAL DESIGN 17 different disease states, such as cognitive dysfunction in major depres- sive disorder and schizophrenia. Nonetheless, it may be more valuable to assume that the dimensions are the same until they are shown to be different. In recognition that drug development in psychiatry is significantly hampered by heterogeneity within diagnostic categories, NIMH estab- lished the RDoC initiative in 2009 to change how patients and non-patients are identified and classified for research purposes, and to encourage re- searchers to think about and study psychopathology in different ways. RDoC became a more dominant paradigm in 2013, when then-Director of NIMH, Thomas Insel, posted a message on his blog, indicating that NIMH was reorienting its research away from DSM categorical syn- dromic diagnoses in favor of an approach that took into account hetero- geneity in diagnosis—essentially the RDoC approach. For example, diagnosing a major depressive episode requires a patient to exhibit five of nine symptoms. Many of these symptoms, such as sleep disruptions or hallucinations, occur across multiple diagnostic categories. With 126 possible combinations of those 9 symptoms, there is substantial hetero- geneity among those diagnosed with depression. Moreover, the occur- rence of symptoms across multiple diagnoses supports the notion that these disorders are not fully distinct, said Sarah Morris, acting head of the NIMH RDoC unit and program chief for schizophrenia spectrum dis- orders research. The primacy of neurobiology over phenomenology is further supported by data from the Psychiatric Genomics Consortium, which showed high rates of shared heritability among diagnoses (Cross- Disorder Group of the Psychiatric Genomics, Consortium, 2013). In an- other example, there is evidence of shared patterns of gray matter loss among patients with schizophrenia, bipolar disorder, depression, addic- tion, obsessive-compulsive disorder, and anxiety (Goodkind et al., 2015). Instead of grouping patients into heterogeneous diagnostic groups, RDoC provides a framework for classifying participants according to neurobehavioral constructs, based on what is known about the brain and behavior, rather than by heterogeneous diagnoses. Morris emphasized that the framework is a hypothesis, which assumes a developmental per- spective from prenatal to late life as well as the ubiquitous impact of the environment across that developmental trajectory. It frames classification in terms of five domains and constructs associated with those domains (see Table 3-1). The list is dynamic and flexible enough to change over 

18 NEUROSCIENCE TRIALS OF THE FUTURE TABLE 3-1 RDoC Domains and Constructs Domains Constructs Cognitive systems Attention Perception Declarative memory Language behavior Cognitive (effortful) control Working memory Positive valence systems Approach motivation Initial responsiveness to reward Sustained responsiveness to reward Reward learning Habit Negative valence systems Acute threat (“fear”) Potential threat (“anxiety”) Sustained threat Loss Frustrative non-reward Arousal and regulatory systems Arousal Circadian rhythms Sleep and wakefulness Systems for social processes Affiliation and attachment Social communication Perception and understanding of self Perception and understanding of others SOURCE: Presented by Sarah Morris at the Workshop on Neuroscience Trials of the Future, March 3, 2016. time. It assumes dimensionality among disorders and between illness and health, as well as interaction among constructs. Constructs can be meas- ured at multiple units of analysis (genes, molecules, cells, circuits, physi- ology, behavior, and self-report), each of which informs and constrains the others. NIMH encourages investigators to adopt RDoC principles in their clinical trials by focusing on functional domains or symptoms, thus in- creasing the probability that participants’ disorders will share the same mechanism. Figure 3-1 illustrates NIMH’s vision of this approach (Insel and Cuthbert, 2015). RDoC also has a database that is integrated with other NIMH data-sharing resources, together providing data from more 

CLINIC CAL TRIAL DESI SIGN 19 than 100,000 subjeccts who are av vailable for hhypothesis tessting or hypotth- esis geeneration. Morris also meentioned that using the RD DoC framewoork will requiire creativ vity with regaard to recruitting a samplee that providees the strongeest test off a hypothesiss, given that the treatmennt focus is onn functional ddo- mains or symptomss rather than unitary u diagnnostic categorries. Enrollmeent criteriaa could be diiagnostically agnostic; an example is rrecruiting bassed on preesence of a syymptom such as psychosiss regardless oof the particullar diagnoosis (e.g., sch hizophrenia orr some other disorder), orr perhaps in tthe future,, expression of a particu ular gene relaated to a psychopathologgic mechaanism. Altern natively, diaggnoses could be used as a proxy forr a sympto om such as pssychoses. FIGUR RE 3-1 RDoC in neurosciencce trials of the ffuture. SOURC CES: Presented by Sarah Mo orris at the Woorkshop on Neuuroscience Trials of the Future, F March 3, 2016. From m Insel, T. R., aand B. N. Cuthbbert. 2015. Braain disordeers? Precisely. Science 348((6234):499–5000. Reprinted with permissiion from American A Assocciation for the Advancement of Science. 

20 NEUROSCIENCE TRIALS OF THE FUTURE BETTER DISEASE MODELS Clinical trials in psychiatry are typically built around the DSM crite- ria, using neuropsychological tests as secondary outcome measures. Bilder suggested that classical psychometric tools (e.g., item response theory and computer adaptive testing) can be used in new ways. He sug- gested going beyond the psychometric model and for the field to consid- er causal models. For example, the classical psychometric explanation of depression assumes that there is an entity, major depression, which caus- es depressive symptoms, weight gain, sleep disturbances, etc. A network or causal modeling approach, by contrast, posits that depression results from the causal interplay among symptoms (Borsboom and Cramer, 2013). It also incorporates the concept of the progression of symptoms over time. Thus, chronic stress leads to depressed mood, which leads to self-reproach, which leads to insomnia, which leads to fatigue, which leads to problems with concentration. Bilder described a network model, built on data from more than 3,000 participants being treated for depres- sion, that identified 28 interconnected nodes, including both DSM and non-DSM symptoms (Fried et al., 2016) (see Figure 3-2). Biological validity is what Bilder called the “holy grail” of drug de- velopment, and viewing the multidimensionality of diseases is one step in that direction. Recent research has begun to demonstrate large genetic correlations where both diagnostic categories and dimensions, or symp- toms within categories, have shared genetic contributions (see Kendler et al., 2011; Lahey et al., 2011). Population-based studies also show great overlap, implying shared genetic plus environmental contributions; this pattern of associations is not limited to psychiatric syndromes. An analy- sis of associations between psychiatric disorders with other medical con- ditions culled from 1.5 million patient records showed an increased risk of nearly every psychiatric disorder with every other medical disorder, suggesting significant shared genetic variation (Rzhetsky et al., 2007). Thus, Bilder advocated for the use of dimensions and clusters of cat- egories as treatment targets, as well as focusing on subgroups not only within disorders, but across syndromal boundaries. He pointed to a recent study investigating the genetic basis of major depressive disorder (MDD) in a population of Han Chinese women. By focusing on a subpopulation of women who had a particularly severe form of melancholic depression, the investigators were able to identify two genetic loci for MDD (Cai et al., 2015). Bilder also pointed to the need for validated biomarkers and expli- cation of intermediate phenotypes to flesh out these cross-level links. 

CLINIC CAL TRIAL DESI SIGN 21 FIGUR RE 3-2 A netw work model of depression. d NOTE:: A: Network co ontaining 28 off the Inventory ffor Depressive S Symptomatologgy- Cliniciaan (IDS-C) depression symp ptoms. Green liines represent positive assoccia- tions, red lines negativ ve ones, and thhe thickness andd brightness off an edge indicaate the assoociation streng gth. The layoutt is based on thhe Fruchtermann–Reingold alggo- rithm thhat places the nodes n with stroonger and/or mmore connectionns closer togethher and thee most central nodes n into the center. c B: Nodee strength centrrality estimates of the 28 IDS-C I depressiion symptoms, including 95 ppercent confidennce intervals. Short codes: c agi = psychomotor p ag gitation; anx = anxious/tensee; app = appettite change; bla = self-blaame/worthless; con = concenntration/decisioons; ene = enerrgy loss; gaas = gastrointeestinal problem ms; hyp = hypeersomnia; in1 = early insomnnia; in2 = mid m insomnia; in3 i = late insom mnia; inp = intnterpersonal sennsitivity; int = in- terest lo oss; irr = irritab bility; pan = paanic/phobia; paar = paralysis; ppes = pessimissm; ple = pleasure loss; qu ua = mood quaality; rea = moood reactivity; reet = psychomotor retardattion; sad = sadn ness; sex = losss of sexual inteerest; som = som matic complainnts; sui = suuicidal ideation n; sym = symp pathetic arousall; var = diurnall variation; weii = weight change. SOURC CES: Presented d by Robert Biilder at the Woorkshop on Neeuroscience Triials of the Future, F March 3, 2016. Repriinted from the Journal of Affffective Disordeers, 189, Frried, E. I., S. Epskamp, R. M.M Nesse, F. Tuerlinckx, annd D. Borsbooom, What are a “good” dep pression sympttoms? Comparring the centrallity of DSM aand non-DS SM symptoms of depression n in a networkk analysis, 3144–320, Copyrigght (2016),, with permissioon from Elsevier. 

22 NEUROSCIENCE TRIALS OF THE FUTURE There has also been a great deal of interest in leveraging and using ge- netic models to facilitate stratification of patients and validate targets, said Perry Nisen. However, he reminded workshop participants that association and causality are not the same; connecting a gene or mutation at a biologi- cal level to a disease process is critical. Similarly, he urged caution in in- terpreting cellular and animal models without a thorough understanding of the underlying biology of the disease. BIOMARKERS Biomarkers, defined as characteristics that can be objectively meas- ured and evaluated as indicators of biological processes (Biomarkers Definitions Working Group, 2001), have become essential across differ- ent phases of drug development. They are typically categorized as diag- nostic, prognostic, predictive, disease progression, or pharmacodynamic, according to their intended use. For example, pharmacodynamic bio- markers may be used in early-stage drug development to demonstrate target engagement, while in later stages, prognostic or predictive biomarkers may be used to stratify participants in a clinical trial or demonstrate treatment response. Throughout the workshop, several participants expressed the need for more validated biomarkers in the field. In a recent systematic review examining published articles on biomarkers in psychosis since 2012, of 3,200 articles, the authors found that most of the biomarkers were diag- nostic, with limited utility for precision medicine studies (Prata et al., 2014). Of the 257 potentially useful prognostic, predictive, or monitoring biomarkers studied, only one had both a reasonable effect size and high quality of evidence, yet even that one had sensitivity of only 21 percent, not enough to change clinical care. According to Shitij Kapur, this study illustrated several problems that plague the biomedical, including the neuroscience, literature: publication bias, lack of replication, or “approx- imate” replication (i.e., replication studies that use different measures and modalities [see Van Snellenberg et al., 2006]); insufficiently pow- ered studies (Button et al., 2013); and insufficient attention to clinical significance (e.g., Levine et al., 2015). Anil Malhotra, director of psychiatry research at Zucker Hillside Hospital and professor of molecular medicine and psychiatry at Hofstra Northwell School of Medicine at Hofstra University, illustrated the need for biomarkers using a recent trial that compared the efficacy of acute 

CLINICAL TRIAL DESIGN 23 treatment with aripiprazole2 and risperidone3 for first episode schizo- phrenia and related disorders (Robinson et al., 2015). The researchers examined the change in psychotic symptoms over 3 months, but despite the relatively large number of participants in the trial (198 patients), no difference was seen. Interestingly, however, there was significant vari- ance in response. If it were possible to identify in advance who was like- ly to benefit from a specific treatment through the use of biomarkers, it might be possible to achieve greater power with smaller studies, said Malhotra. He described some of the advantages and disadvantages of genetic and neuroimaging biomarkers for psychiatric disorders, noting, however, that there are many other types of biomarkers, including biochemical markers from the cerebrospinal fluid (CSF) and plasma, cognitive and neuropsychological measures, among others. With regard to plasma me- tabolites, although they are popular because of the relatively easy access, so far it has been very difficult to extrapolate from the plasma to the brain, said Malhotra. Advantages of genetic biomarkers include the ease of access to DNA from blood or saliva and the stability of genotype over time, mean- ing you can collect samples after a trial for a prospective analysis. Dis- advantages primarily relate to power: the sample sizes of most genetic studies are quite large. Neuroimaging biomarkers also have both ad- vantages and disadvantages, said Malhotra. The most notable advantage is the ability to directly assess brain structure and function, including the ability to assess specific regions and circuits within the brain. Disad- vantages include the difficulty of precise replication, potential confounds from environmental factors such as prior treatment, and subject ac- ceptance. Moreover, it is not always clear what is being measured, he said. Many different neuroimaging modalities may yield potential bio- markers, including structural magnetic resonance imaging (MRI), diffu- sion tensor imaging (DTI), task-based functional MRI (fMRI), and resting-state fMRI (see Table 3-2). Resting-state fMRI has shown particular promise in providing a good signal and also is one of the easier measures to access, said Malhotra. It provides a measure of brain activity in the absence of an externally prompted task as a means of defining functional networks. It also enables 2 Aripiprazole is an atypical antipsychotic, used primarily to treat schizophrenia and bi- polar disorder. 3 Risperidone is an antipsychotic used primarily to treat schizophrenia, bipolar disorder, and behavior problems. 

24 NEUROSCIENCE TRIALS OF THE FUTURE TABLE 3-2 Neuroimaging Approaches to the Heterogeneity of Anti- psychotic Response Modality Characteristic Measured Structural CT or MRI Assessment of morphology Diffusion tensor imaging Putative measure of white-matter integrity Task-based fMRI Change in BOLD signal during conduct of a cognitive/behavioral/emotional activation task Resting-state fMRI Correlation in BOLD signal during “rest” NOTE: BOLD signal = blood-oxygen-level‒dependent changes, induced by blood flow; CT = computerized tomography; MRI = magnetic resonance imaging. SOURCE: Presented by Anil Malhotra at the Workshop on Neuroscience Trials of the Future, March 3, 2016. investigators to assess the functional connectivity among multiple re- gions of the brain. For example, Malhotra and colleagues showed that improvement in psychotic symptoms over 12 weeks in individuals with first episode schizophrenia correlated with increased connectivity in cor- tical striatal circuits, suggesting that increasing connectivity in these cir- cuits could be used as a biomarker of antipsychotic efficacy (Sarpal et al., 2015). Indeed, in a more recent study, Malhotra’s team used baseline striatal functional connectivity to predict antipsychotic drug response (Sarpal et al., 2016). The striatal connectivity index4 (SCI) has also been applied to other clinical studies, including the aripiprazole and risperidone study men- tioned at the beginning of this chapter, in which no difference in efficacy was seen. Among study participants who had undergone baseline resting- state fMRI scans, responders tended to have lower SCI than non- responders, further supporting its potential as a predictive biomarker. Several participants highlighted that biomarkers for nervous system disorders are somewhat more developed in neurology than psychiatry. For example, Story Landis, former director of the National Institute of Neurological Disorders and Stroke (NINDS), commented that the identi- fication of neuroimaging as a biomarker for immunological disturbances transformed drug development for multiple sclerosis (MS), where there are now drugs that slow progression of the disease. However, much re- mains to be learned with regard to how and why these drugs work as well as on other aspects of MS, such as the neurodegenerative component. 4 An aggregate of the 91 couplets of connectivity between the striatum and the cortex, was used by the researchers to determine antipsychotic drug efficacy. 

CLINICAL TRIAL DESIGN 25 There has been a particularly substantial effort to develop biomarkers for Alzheimer’s disease (AD), exemplified by ADNI. ADNI was launched in 2004 with the initial goal of developing imaging and bio- chemical biomarkers for the early detection of AD, as well as for use in clinical trials. However, the impact of ADNI on the AD field and beyond has been far broader, said Alice Chen-Plotkin, assistant professor of neu- rology at the Perelman School of Medicine, University of Pennsylvania. It set a precedent for open data sharing in the neurodegenerative disease space and demonstrated the potential benefits of bringing together ex- perts from industry, governmental agencies, academia, and private foun- dations to provide the funds required to address a monumental problem (Weiner et al., 2015). The original ADNI cohort consisted of 200 normal controls, 200 people with overt AD, and 400 people with amnestic mild cognitive im- pairment (aMCI), which was thought to possibly represent prodromal AD. Study participants were assessed clinically and with biomarker stud- ies at 6-month intervals for 4 years. Following the initial 5-year study, additional funding was obtained in 2009 with a Grand Opportunities grant (ADNI-GO) and in 2011 with a renewal (ADNI-2), enabling an additional 550 people to be enrolled and the focus to shift to earlier phas- es of the disease. Since its inception, data from ADNI have resulted in more than 600 publications (Weiner et al., 2015) that have transformed the understand- ing of AD. Chen-Plotkin summarized some of the learnings from ADNI: • Biochemical biomarkers in the CSF can discriminate individuals with aMCI who will go on to develop AD from those who will not (Shaw et al., 2009). • Positron emission tomography (PET) scans using ligands that bind to beta amyloid (Aβ, the protein found in the amyloid plaques seen at autopsy) can demonstrate the deposition of amy- loid in the brain in vivo (Clark et al., 2011). • These measurements are meaningful if one knows the quantita- tion is accurate and can be reproduced in any lab that follows the standardized procedure. ADNI has already led to changes in how AD trials are conducted, said Chen-Plotkin. Phase III trials are being conducted in preclinical stages of AD using amyloid imaging and CSF biomarkers as entry crite- ria to enrich for patients on the AD trajectory. Biomarkers are also cur- 

26 NEUROSCIENCE TRIALS OF THE FUTURE rently being incorporated into diagnostic criteria in the research settings, but may be used for clinical diagnosis in the future (Sperling et al., 2011). Identifying biomarkers to expedite clinical trials in Parkinson’s dis- ease (PD) has proved to be more difficult, said Chen-Plotkin. Although both AD and PD are characterized by misfolding and aggregation of a central culprit molecule—Aβ in AD and α-synuclein in PD—no specific imaging ligand for α-synuclein has been identified, and levels of α- synuclein in CSF are not predictive of disease progression (Kang et al., 2013). Chen-Plotkin, along with other PD researchers, think an unbiased screening approach, made possible by advances in genomics and prote- omics, will be needed to identify predictive biomarkers for PD (Chen- Plotkin, 2014). The PD community (e.g., researchers and patient advoca- cy groups), like the AD community before it, now is organizing around a pipeline for discovery, replication, and further development of novel PD biomarkers. In 2012, NINDS launched the Parkinson’s Disease Bi- omarker Program (PDBP), which has collected biospecimens and clinical data from more than 1,000 people, stored them in a central repository, and made them available for discovery efforts by the neuroscience re- search community (Rosenthal et al., 2015). Meanwhile, the Parkinson’s Progression Markers Initiative (PPMI), sponsored by the Michael J. Fox Foundation for Parkinson’s Research, has collected samples along with clinical and behavioral assessments from multiple cohorts, including normal controls and individuals with PD, as well as those in the prodro- mal stage of PD. These samples and data are available to the research community. Chen-Plotkin’s group at the University of Pennsylvania was one of the first to use this pipeline. In 2013, they identified a candidate protein biomarker for PD using an unbiased screening approach on samples ac- quired at the university. They showed that higher plasma levels of Apolipoprotein A1 (ApoA1) were correlated with older age at PD onset and less severe PD (Qiang et al., 2013). Then, using samples from PPMI, they replicated the study, thus providing the first report of a plasma- based biomarker of disease progression in PD and suggesting that ApoA1 may represent a therapeutic target (Swanson et al., 2015). Chen-Plotkin and colleagues were also interested in replicating a study that suggested that low levels of epidermal growth factor in the blood were predictive of which PD patients would become demented. However, they observed substantial site-to-site variability in the meas- urement of epidermal growth factor at PPMI sites, highlighting the need for strict adherence to standardized protocols. 

CLINICAL TRIAL DESIGN 27 Moreover, there are gaps in our knowledge about biomarkers, which need to be overcome, said Bilder. For example, in a study of the associa- tion of amyloid burden with disruption of the default mode network, nei- ther of these measures were associated with cognitive impairment (Hedden et al., 2009). In addition, several participants noted that there is a great need for biomarkers with predictive power. This will likely re- quire a multimodal approach rather than a single determinant biomarker, said Malhotra. Luca Pani, director general of the Italian Medicines Agency (AIFA), said biomarkers are also needed to show that a treat- ment makes sense in terms of value. For example, if a biomarker could predict that a specific treatment for hepatitis C reduces the need for transplantation in some patients, the savings could influence the approval decision. The predictive biomarkers (diagnostic) in oncology play a criti- cal role in understanding molecular and cellular mechanisms, which drive tumor initiation, maintenance and progression. They help to opti- mize therapy decisions, as they provide information on the likelihood of response to a given chemotherapeutic. However, he noted that this would be more difficult for psychiatry and neurology area. Chen-Plotkin and S. Kapur emphasized the need for building interna- tional consortia and public–private partnerships to advance the develop- ment of new tools. For example, an international consortium called Novel Methods leading to New Medications in Depression and Schizo- phrenia (NEWMEDS) has developed a tool called a clinical significance calculator to help investigators estimate whether a predictive biomarker for depression is likely to demonstrate clinical significance, shifting at- tention away from a focus on statistical significance, or p-values (Uher et al., 2012), said S. Kapur. In addition, Chen-Plotkin noted that bioreposi- tories and sharing of biospecimens will enable the more efficient devel- opment of a wide range of biomarkers. CLINICALLY MEANINGFUL OUTCOMES In recent years, regulators and payers have increasingly required that clinical studies demonstrate not only statistical significance of an effect, but even more importantly, clinical significance (Ranganathan et al., 2015). This view was also reflected in comments from many workshop participants. The bottom line, said Romano, is that no payer will provide coverage for a product that does not show clinical relevance. Studies need to show that a difference is relevant not against placebo, but against 

28 NEUROSCIENCE TRIALS OF THE FUTURE comparators, he added. S. Kapur noted that effect size is not the same as clinical significance. He further proposed developing a framework for assessing the clinical significance of potential innovative options that will make health care cheaper, not just “flashier” and stratified. Clinically relevant endpoints are also important to promote clinical implementation. Publication in an appropriate peer-reviewed journal provides an avenue for acceptance by the community, yet an analysis of randomized clinical trials supported by the National Heart, Lung, and Blood Institute showed that only 57 percent of studies were published within 30 months of completion of the trial. However, those trials with clinical endpoints were published significantly sooner than those with surrogate endpoints (Gordon et al., 2013), noted Adrian Felipe Hernandez, professor of medicine at the Duke University School of Medicine. NOVEL CLINICAL TRIAL DESIGNS Statistical Approaches and Considerations Many design and methodology approaches can increase trial effi- ciency, said Michael Pencina, director of biostatistics at the Duke Clini- cal Research Institute and professor of biostatistics and bioinformatics at Duke University. In general, he said, regulatory bodies are most open to innovative clinical trial designs in smaller studies and in earlier stages of drug development. The design has to be fit-for-purpose, and appropriate statistical expertise is needed in the planning stage. He gave a brief over- view of the various design choices: • Event-driven trials are appropriate if an outcome can be meas- ured over time and the duration of the trial is sufficient to enable an adequate number of events to be observed. In contrast to more traditional studies where participants are followed for a specific period of time, time-to-event studies can be stopped when an ad- equate number of outcomes are achieved, and all follow-up is in- cluded in the analysis. Efficiency can be further maximized by blinded interim monitoring of event counts, followed by increas- ing or decreasing the planned study duration, or increasing the sample size. • Composite outcomes, used extensively in cardiology studies, may also increase the efficiency of a study, particularly if there 

CLINICAL TRIAL DESIGN 29 are multiple outcomes that are roughly similar in severity. The composite itself may provide increased power, while analysis of individual components may bring added value. Hierarchical composites combine outcomes of varying severity. In this case, the most severe outcome (e.g., death) is used as the first compar- ator; if this outcome fails to materialize, the next most severe outcome (e.g., hospitalization) could be analyzed, etc. Continu- ous outcomes make this method more powerful. • Adaptive designs, in which a key study feature, such as sample size, duration, or number of treatment arms, can be modified and adapted based on an interim analysis. One example is a “drop- the-losers” design, where treatment arms are dropped if they fail to meet a prescribed threshold, while those that meet this thresh- old are advanced to the next stage. • Enrichment designs typically use biomarkers to ensure a more homogeneous and likely-to-respond study population. Adaptive enrichment designs adapt based on a biomarker. For example, different arms could include participants with different genetic markers. Following interim analysis those arms showing greater effect sizes could be enriched. • Sequential parallel comparison designs are used to reduce the impact of the placebo response. In the first stage of such a de- sign, participants may be randomized to receive treatment or pla- cebo; in the second stage, placebo non-responders are re- randomized. The final summary statistic in this type of trial is then based on a weighted combination of effects from the two stages. • Controls may sometimes be borrowed from historical infor- mation, using several different methods. The simplest, yet con- troversial, approach is to pool historical information with randomized controls. If there is enough homogeneity among the studies, this can substantially reduce sample size. Historical in- formation may also be used to define performance criterion by enabling derivation of an estimated event rate that the treatment being tested must exceed. Another more sophisticated approach involves testing first to see if controls are sufficiently similar for pooling. Even more sophisticated approaches allow investigators to model variations between current and historical data to enable their use. 

30 NEUROSCIENCE TRIALS OF THE FUTURE 2×2 Blind Trial Design Erik Snowberg, professor of economics and political science at the California Institute of Technology, proposed a novel design, which he developed with his colleague Sylvain Chassang, professor of economics and public affairs at Princeton University. Their design leverages the ob- servation that dropouts decline when participants have a greater likeli- hood of receiving treatment, and capitalizes on the “placebo effect.” In the biomedical world, placebo implies a sham treatment; however, econ- omists equate placebo effects with behaviors, which can interact with treatment and affect the efficacy of a treatment. For example, participants who believe they are receiving treatment for depression may change their behavior with the thought that because they are being treated, they will be able to more successfully navigate social situations. Interacting with others may then help alleviate their depression, and the behavior thus influences the treatment effect. Clinical trials that allocate participants equally to the treatment and placebo arms are typically suboptimally powered, said Snowberg, in part because of the high rate of dropouts. More power can be achieved by randomizing more participants to the treatment arm. This not only reduc- es dropouts, but changes behavior, because participants believe they have a higher likelihood of receiving treatment. Moreover, it allows investiga- tors to assess the benefit from the interaction of behavior and treatment. Snowberg and Chassang tested this theory using participant-level data collected by Fournier and colleagues (2010) in a meta-analysis of double- blind RCTs for depression. Data from six trials (three trials each for imi- pramine5 and paroxetine6) were selected. Among the imipramine trials, two allocated participants equally and one allocated 70 percent to receive treatment. Among the paroxetine trials, one allocated equally and the other two allocated 65 percent and 67 percent to receive treatment. Their analysis showed that treatment probability affected participants’ decision to drop out of a trial, and that for paroxetine, but not imipramine, there was an interaction between treatment and behavior that resulted in an improved treatment effect (Chassang et al., 2015). With the 2×2 blind trial that Snowberg and Chassang propose (see Figure 3-3), the aggregate probability of treatment is 50 percent, but the 5 Imipramine is a tricyclic antidepressant used primarily to treat major depressive disorder. 6 Paroxetine is a selective serotonin reuptake inhibitor (SSRI) used to treat major de- pressive disorder and other psychiatric disorders. 

CLINIC CAL TRIAL DESI SIGN 31 FIGUR RE 3-3 2×2 bliind trial design n. SOURC CES: Presented by Erik Snow wberg at the W Workshop on N Neuroscience T Tri- als of th he Future, Marrch 3, 2016, fro om Chassang eet al., 2015. differeent groups allow the investtigator to deccompose the ttreatment effeect into th hree parts: thee effect of beehavior, the eeffect of the ttreatment itseelf, and thhe interaction of behavior and treatmennt. Participannts are first raan- domized into two groups: a hig gh probabilityy (75 percennt) of treatmeent group and a low probability (25 percent) of tr treatment grouup. Participannts are infformed of theeir probabilityy of treatmennt, and the trials are run inn a blinded fashion in the usual waay, with a coombined analyysis of the tw wo groupss. Snowberg said this app proach in Phaase II providees investigatoors with innformation thhat will enablle them to coonduct more ooptimally pow w- ered Phase III studiies. Thee Established d Status Epileepticus Triall Jaiideep Kapur, Eugene Mey yer III professsor of neurosccience and neeu- rologyy at the University of Virg ginia School of Medicinee, described aan- other novel n trial deesign that is being used tto identify thhe best way of treatin ng benzodiazepine-refracto ory status eppilepticus. T The Establishhed Status Epilepticus Trial T (ESETT T) was designned to identifyy the best treaat- ment for f the 35 to 45 4 percent off patients whoo do not respoond to benzoddi- azepinnes, the standaard first line agent for treaatment of staatus epilepticuus. There has been a laack of well-ccontrolled stuudies for this indication, and ment practices vary. In the United Statees, the most ccommonly ussed treatm drug iss fosphenytoiin; however, a newer drugg, levetiracetaam, is easier to use annd has fewer side s y evidence ffrom uncontrrolled studies in effects, yet Europee, India, and other o places suggest s that vvalproic acid iis superior, saaid J. Kappur. The ESE ETT investigaators have theerefore designned a comparra- tive effficacy trial to o determine which w drug is bbest (Bleck ett al., 2013). 

32 NEUROSCIENCE TRIALS OF THE FUTURE The trial, which is being conducted at 51 sites, has several unique characteristics. Recruitment is conducted by emergency department phy- sicians from patients who are transported to the study sites and received benzodiazepines en route by emergency medical personnel. If a patient continues to seize and meets the inclusion criteria, the physician adminis- ters the study drug, which has been provided in prerandomized, blinded study boxes. The outcome is absence of seizures and regaining of con- sciousness at 60 minutes. J. Kapur noted that randomization will also be stratified into three age groups: 2–18 years, 19–65 years, and 66 years or older. The second important feature of ESETT is a Bayesian adaptive de- sign. In designing this trial, investigators ran thousands of simulations based on several different scenarios (e.g., different effect sizes and false- positive rates for the three compounds), which enabled them to select the optimal operating characteristics (e.g., timing of analyses, power, sample size, etc.) with adequate power and a minimum sample size (Connor et al., 2013). The design they selected will start by first enrolling 300 pa- tients, allocated equally to each of the three drugs. After an interim anal- ysis, allocation ratios will be modified based on performance, such that the best performing drug will get more patients and the worst performing drug will get fewer. This type of design ensures that the most effective drug will be given to the largest proportion of patients. J. Kapur noted that the trial will be stopped early for efficacy (one treatment clearly bet- ter) or futility (either all arms are bad or the trial appears unlikely to identify a best and worst treatment). He added that it will be finished and considered a success when the probability that one treatment is the most effective exceeds 0.975. The trial has been funded for 795 patients, but the investigators hope to have an answer and finish early. LESSONS LEARNED FROM OTHER THERAPEUTIC AREAS Successful drug development in other disease areas such as oncology and cardiology may provide insight for neurology and psychiatry. For example, a major advance in the treatment of melanoma required the de- velopment of a targeted regimen for melanomas resistant to therapy. This regimen uses two drugs that inhibit different cell-signaling molecules, and was made possible by leveraging genetic data to facilitate the stratifi- cation of patients, which enabled validation of targets, according to Nisen. Researchers had first discovered the mechanisms of resistance to one 

CLINICAL TRIAL DESIGN 33 type of inhibitor and then proceeded to design a study that demonstrated the efficacy of combined treatment with drugs that inhibit two different cell-signaling molecules (Flaherty et al., 2012; Wagle et al., 2011). One of the important lessons from this effort, said Nisen, is that in addition to having a validated mechanism and a way of assessing an early signal, combining therapeutics at the earliest stages of drug develop- ment should be considered for disorders that are heterogeneous and multi- factorial. In addition, while the conventional wisdom is that studies should be kept as simple as possible, Nisen said this study was extremely complicated and very adaptive, but enabled the investigators to answer fundamental questions about dose, drug‒drug interactions, pharmacoki- netics, safety and tolerability, and clinical activity. Another important lesson from oncology relates to the establishment of Comprehensive Cancer Centers, which provide improved integration between basic science and the clinical enterprise. According to Nisen, about 90 percent of children with cancer participate in clinical trials at these centers, with high cure rates for certain types of cancer (e.g., most childhood acute lymphoblasic leukemia is cured). Adults with cancer are much less likely to be treated at one of these centers, and less than 5 per- cent participate in trials. The survival rate for all adults is much lower. Cardiology provides additional lessons applicable to neuroscience trials of the future. The cardiology field has a relatively strong evidence base upon which treatment guidelines have been developed. Nonetheless, a cultural demand there remains within the cardiology community to fill in the knowledge gaps, according to Hernandez. Thus, networks of car- diologists have collaborated to develop large, multicenter randomized studies. For example, the Global Utilization of Streptokinase and t-PA for Occluded Coronary Arteries (GUSTO) study enrolled more than 40,000 patients at more than 1,000 hospitals in 15 countries despite the required completion of a three-page case report form for each patient (GUSTO Investigators, 1993). Key to harnessing the community’s inter- est is asking a relevant question, said Hernandez. For example, the drug neseritide was approved in 2001 for acute heart failure with a pivotal study of fewer than 500 patients. Widespread use, however, indicated that the drug was associated with acute adverse effects. This led to the creation of the Acute Study of Clinical Effectiveness of Nesiritide and Decompensated Heart Failure (Hernandez et al., 2009), a large pragmatic trial focused on understanding clinically meaningful outcomes in the context of real-world use. Despite the negative press about nesiritide, the cardiology community came together and successfully enrolled more 

34 NEUROSCIENCE TRIALS OF THE FUTURE than 3,000 participants, exceeding its projected enrollment. Now, the cardiology community is moving forward with a large-scale (20,000 par- ticipants), pragmatic, adaptable, patient-centric, randomized controlled trial (RCT) using the National Patient-Centered Clinical Research Net- work (PCORnet). This study will leverage data from electronic health records and collect patient-reported information electronically through patient portals. Another lesson learned from cardiology studies is the importance of making sure the background therapy is relevant to the market. For exam- ple, in a study of ticagrelor,7 outcomes in North America were substan- tially poorer than those in other parts of the world (Wallentin et al., 2009). A post-hoc study identified the reason: A different dose of aspirin being used as an antiplatelet background therapy, said Hernandez. He added that selecting the appropriate outcome measure for the trial is also important. In addition, a new paradigm that has emerged from cardiology is the registry-based randomized clinical trial (Lauer and D’Agostino, 2013), which the authors of the study called “the next disruptive technology in clinical research,” said Hernandez. This approach leveraged clinical in- formation collected by a preexisting observational registry to identify potential participants, which markedly accelerated enrollment and elimi- nated the need for lengthy case report forms (CRFs). Pencina noted that electronic health records and registry-enabled trials may also be useful in allowing investigators to run very large, simple trials for a fraction of the cost of more typical trials. Finally, a few participants noted that value of trials in which patients serve as their control (also known as N of 1). Although not frequently used for trials for nervous system disorders, N of 1 trials may lead to more individualized treatment (Lillie et al., 2011). DEVELOPING MORE EFFECTIVE THERAPEUTICS THROUGH PRECISION MEDICINE: IMPLICATIONS FOR CLINICAL TRIALS Opportunities may also emerge from examination of success in other fields, said Shitij Kapur, executive dean and head of school at the Insti- tute of Psychiatry, Psychology & Neuroscience, King’s College London. In oncology, precision medicine has emerged as a valuable and innova- 7 Ticagrelor is a platelet aggregate inhibitor used to prevent strokes and heart attacks. 

CLINICAL TRIAL DESIGN 35 tive approach for treating some of the most intractable types of cancer; many of the advances are found in the treatment of breast cancer. Today, nearly all newly diagnosed breast cancers are tested for the presence of the human epidermal growth factor receptor (HER2/neu). These efforts in the breast cancer field highlight the need for patience and persistence, said S. Kapur. Two recent papers show that following the principles of personalized medicine double the chances of success (Cook et al., 2014; Nelson et al., 2015). However, the movement toward personalized medicine has not yet resulted in precision medicine trials in nervous system disorders despite a few promising leads (Liu et al., 2012; Volpi et al., 2009). S. Kapur predicted that precision medicine in neurology and psychia- try will enrich and modify rather than replace current practice, presenting both challenges and opportunities for clinical trial design. Specifically, he noted that developing biomarkers that predict disease progression and treatment response longitudinally is needed for nervous system disorders. Several participants added that using multimodal approaches rather than a single approach may help to show which drugs or other intervention approaches work best for which patients. In addition, using big data ap- proaches may help to identify correlations among diverse types of patient data and outcomes. Robert Califf emphasized that recruiting as many volunteers as possible to participate in research as a normal part of their patient care, through the Precision Medicine Initiative, will be important for therapeutic development (e.g., Slamon et al., 1987; Smith et al., 2007).