Age of onset in genetic prion disease and the design of preventive clinical trials

Regulatory agencies worldwide have adopted programs to facilitate drug development for diseases where the traditional approach of a randomized trial with a clinical endpoint is expected to be prohibitively lengthy or difficult. Here we provide quantitative evidence that this criterion is met for the prevention of genetic prion disease. We assemble age of onset or death data from N=1,094 individuals with high penetrance mutations in the prion protein gene (PRNP), generate survival and hazard curves, and estimate statistical power for clinical trials. We show that, due to dramatic and unexplained variability in age of onset, randomized preventive trials would require hundreds or thousands of at-risk individuals in order to be statistically powered for an endpoint of clinical onset, posing prohibitive cost and delay and likely exceeding the number of individuals available for such trials. Instead, the characterization of biomarkers suitable to serve as surrogate endpoints will be essential for the prevention of genetic prion disease. Biomarker-based trials may require post-marketing studies to confirm clinical benefit. Parameters such as longer trial duration, increased enrollment, and the use of historical controls in a post-marketing study could provide opportunities for subsequent determination of clinical benefit.


Introduction
Placebo-controlled, double-blind, randomized trials with a clinical endpoint -a measure of how patients feel or function -constitute the gold standard for demonstration of therapeutic efficacy and, where feasible, are strongly preferred for approval of new drugs. Regulators worldwide have recognized, however, that in some diseases the duration of such trials may unduly delay patient access to potentially life-saving drugs. Many agencies have therefore created programs to support drug development in this situation. For instance, the United States Food and Drug Administration (FDA) Accelerated Approval program 1 provides for conditional approval based on trials using surrogate endpoints, including biomarkers, with a requirement for post-marketing studies to confirm clinical benefit 2 . Honoring the specifics of each disease, FDA "consider[s] how to incorporate novel approaches into the review of surrogate endpoints… especially in instances where the low prevalence of a disease renders the existence or collection of other types of data unlikely or impractical" 3 . Here we present evidence that genetic prion disease meets this criterion.
Prion disease is a fatal and, at present, incurable neurodegenerative disease caused by the misfolding of the prion protein, PrP, encoded by the gene PRNP 4 . Most subtypes of prion disease are extremely rapid, leading from first symptom to death in several months 5 . Prion diseases are transmissible, but today few cases are known to be acquired by infection. ~85% of prion disease cases are termed "sporadic," meaning they arise spontaneously with no known environmental or genetic trigger, while ~15% of cases possess protein-altering rare variants in PRNP, a subset of which are highly penetrant 6 . Various genetic subtypes of prion disease include fatal familial insomnia, genetic Creutzfeldt-Jakob disease, and Gertsmann-Sträussler-Scheinker disease.
To date, all completed clinical trials in prion disease have recruited only symptomatic patients, mostly with sporadic prion disease, and have used cognitive, functional, or survival endpoints [7][8][9][10][11][12][13][14] . By the time of diagnosis many prion disease patients are in a state of advanced dementia, and even a therapy that halted the disease process entirely at this stage might only preserve the patient in a state with little or no quality of life 15 . Moreover, preclinical proofs of concept argue that a preventive, rather than therapeutic, approach is more likely to be effective. Multiple antiprion agents have been discovered that extend the survival time of prion-infected mice by 2-4X when administered long before symptoms, yet these have diminished effects at later timepoints, and no effect when administered after clinical onset [16][17][18][19] . These observations indicate a need to enable preventive trials in presymptomatic individuals at risk for genetic prion disease.
The ongoing preventive trial of crenezumab, an anti-amyloid β antibody, for PSEN1 E280A early-onset Alzheimer's disease, follows a design where presymptomatic individuals are randomized to drug or placebo and followed for five years to a cognitive endpoint 20 . While this represents one model for preventive trials in neurodegeneration, we hypothesized that this approach might be challenging for genetic prion disease due to its variable age of onset 21,22 , small presymptomatic patient population 23 , and more limited financial incentives for pharmaceutical companies. To test this hypothesis, we set out to aggregate age of onset data in genetic prion disease, generate survival and hazard curves, and simulate statistical power for randomized pre-approval trials with a clinical endpoint. We also set out to investigate the feasibility of one potential alternative: post-marketing studies using historical controls to confirm clinical benefit, following Accelerated Approval on a surrogate biomarker endpoint.

Age of onset in genetic prion disease
We reasoned that any preventive trial with a clinical endpoint in genetic prion disease would derive most of its statistical power from individuals with high penetrance PRNP variants. Some PRNP variants can be identified as highly penetrant by their extreme enrichment in cases over population controls, but many variants are too rare in both groups for meaningful comparison 6 . We therefore reviewed the literature on 69 reportedly pathogenic PRNP variants and identified 27 variants for which Mendelian segregation has been reported in at least one family with at least three affected individuals and/or for which a de novo mutation in a case has been identified (Table S1), thus suggesting high penetrance.
We examined the frequency of these putative high penetrance variants in a recent case series 6 . The top three variants -E200K, P102L, and D178N -collectively explain 85% of high penetrance cases ( Figure S1). Each of these arises from a CpG transition (a C to T DNA change where the adjacent base is G), a type of variant which occurs by spontaneous mutation 10-100X more often than other mutation types 24,25 , explaining the recurrence of these three mutations on multiple PRNP haplotypes in families worldwide 6,26,27 . Therefore, regardless of the population studied, these three variants are likely to account for a large fraction of genetic prion disease cases with high penetrance variants. For this reason, we focused our analysis primarily on individuals with these three variants. We aggregated age of onset and/or age of death data on N=1,001 individuals with the E200K, P102L, or D178N mutations from nine study centers worldwide (Table 1 and Table S2), encompassing both direct clinical reports and family histories (see Methods), and including censored individuals. Statistics on N=93 individuals with the next four mutations most common in cases -5-OPRI (insertion of five extra octapeptide repeats), 6-OPRI, P105L, and A117V -are included in Table S3. We used these data to compile life tables and computed the annual hazard -risk of onset in each year of life -for each mutation (Supplementary Life Tables).
We found wide variability in age of onset (Table 1), consistent with previous reports 21,22,28 . An implication of this variability is that high lifetime risk arises not from certain onset at a specific age, but from modest risk in any given year of life, accumulated over many decades of exposure. This poses a challenge for following presymptomatic individuals to onset in a preventive clinical trial, as it is difficult to ascertain a group of individuals for whom onset is imminent. For example, even at age 57, an E200K individual has only a 5% probability of disease onset occurring in any given year. This means that 20 person-years of follow-up for E200K individuals around this age would be expected to result in only one observed disease onset. Annual hazards do rise with age, but as they reach high levels, the number of surviving individuals also dwindles ( Figure 1). For the three most common mutations, the annual hazard remains below 10% until after the majority of people have already died ( Figure 1, Figure S3, and Supplementary Life Tables). Similarly, the median number of years until onset, conditioned on an individual's current age, remains ≥5 years until after the median age of onset has passed (Supplementary Life Tables). The next four most common mutations have tighter age of onset distributions, and so reach higher annual hazards sooner ( Figure S3), but these mutations are also much rarer, accounting for only 10% of cases with a high penetrance variant 6 Figure S3.

Power for randomized pre-approval trials with a clinical endpoint
We set out to calculate how many individuals would need to enroll in order to power prevention trials with an endpoint of disease onset, using the calculated age-dependent hazards for each mutation. While younger individuals or those with a mutation of modest penetrance might seek to enroll in trials or take a preventive drug, they would not contribute much statistical power to an endpoint of clinical onset. We therefore chose to base our power calculations on individuals with the three most common high penetrance mutations between age 40 and 80.
We estimated how many individuals in this age range have high penetrance PRNP mutations. It is estimated based on disease prevalence that 1-2 people per 100,000 in the general population harbor high penetrance PRNP mutations (ref. 6 and Supplementary Discussion), but at present, many remain unaware of their risk due to underdiagnosis 29 of affected family members, and few choose predictive testing 23  number of positive predictive genetic test results that have been provided in the U.S. is N=221 (Supplementary Discussion), and based on the estimated proportion of high penetrance variants 6 (75%), and the estimated proportion of positive test result recipients 23 over age 40 (36%) we estimate there are currently ~60 people in the U.S. who are age 40 or older and hold a positive predictive test for a highly penetrant variant.
We used published formulae 30 (see Methods) to calculate statistical power for a log-rank survival test in randomized clinical trials (Table 2). Across the three mutations and weighted by their prevalence among cases ( Figure S1) and number of surviving individuals at each age (Figure 1), the average annual probability of onset for individuals aged 40 to 80 is 4.6%. We used the 4.6% figure as a baseline hazard, and made the following assumptions: presymptomatic individuals are randomized half to drug and half to placebo and followed for 5 years with an endpoint of clinical onset; events in the first year are ignored as a "run-in" period to ensure sufficient drug exposure among individuals analyzed; the withdrawal rate is 15.2% annually (the median value from eight prevention trials reviewed, Table S4); and the trial is designed for 80% power at the P=0.05 threshold. We then performed power calculations for such a trial as a function of the hazard ratio -the ratio of annual risk of onset in drug-treated individuals to that in placebo-treated individuals. For context, we also determined the effect size, in median years of healthy life added, to which each hazard ratio corresponds ( Table 2). The calculations are sensitive to which mutations are included, the "run-in" period, the number of years of follow-up, and the assumed withdrawal rate, but we explored a range of different assumptions and none support a different overall interpretation of the data (Table S5 and Discussion). In particular, the assumption of a 15.2% annual withdrawal rate means that only 44% of original participants remain after 5 years, but even reducing the withdrawal rate to zero only lowers the numbers of participants required by one-third (Table S5). Because FDA has cautioned against rare disease trial designs that assume a large effect size 31   The above power calculations simplistically assume a uniform baseline hazard across all participants, regardless of age and PRNP mutation. We also used a simulation to account for the full shape of the hazard curve and diversity of genetic mutations, but the simulated power results were similar to those in Table 2 (see Table S6). Stratification by PRNP mutation did not improve power in our simulations (Supplementary Discussion), perhaps because age of onset distributions (Table 1) are wide and overlapping, such that PRNP mutation explains only a minority of the overall variance in age of onset (adjusted R^2 = 0.15, linear regression, P < 1e-32).
Statistical power might be improved by stratifying clinical trial analysis by relevant additional variables, but there are currently no variables that help to explain age of onset (Supplementary Discussion, Table S7, and Figure S4 -5). For instance, we found no sex effect, and no evidence that parent and child age of onset are correlated after controlling for PRNP mutation and for child's year of birth, a variable that captures some effects of ascertainment bias 32 (Table S7). A common genetic variant, PRNP M129V, is known to affect the clinical and pathological presentation of many forms of prion disease 33 as well as the risk of sporadic and acquired prion disease 34 . This variant has previously been reported to affect age of onset in some forms of genetic prion disease but not others 21,32,35,36 . We found no evidence that codon 129 affects age of onset for P102L or E200K individuals (Table S7 and Figure S4). For D178N, our data are suggestive that a 129VV genotype may predispose to earlier onset than MM or MV genotypes ( Figure S4 and Supplementary Discussion), but in the overall dataset, codon 129 failed to explain additional variance in age of onset (Supplementary Discussion).
Based on this analysis, at present it is not possible to adequately power a randomized preapproval prevention trial with an endpoint of clinical onset in genetic prion disease. For example, for a drug that reduces annual risk by half (hazard ratio of 0.5), powering such a trial would require 813 participants age 40 or older, and even for a drug that reduces annual risk by ten-fold (hazard ratio of 0.1), 101 participants would be required (Table 2), versus the ~60 currently estimated to exist in the U.S. Key assumptions underpinning this analysis may change with time: new stratifying variables could help to predict age of onset, or a first drug for prion disease could improve diagnosis and recruitment (see Discussion). However, the insight that randomized pre-approval prevention trials with a clinical endpoint may not be feasible today has implications for drug development efforts likely to reach the clinic while current assumptions hold. For this reason, we next turned our attention to the possibility that a preventive drug might be developed through the Accelerated Approval pathway using a surrogate biomarker endpoint.

Power for post-marketing studies
We asked whether, if Accelerated Approval were achieved, the required post-marketing studies to confirm clinical benefit could be adequately powered by following drug-treated individuals to clinical onset and comparing their survival to that of historical controls. Such a trial design could increase power but also introduce bias; we considered each issue in turn.
We identified several factors that are likely to decrease the number of participants required to power such a study compared to its randomized pre-approval equivalent: all, rather than half, of individuals are drug-treated; the number of historical controls can be large; a longer trial duration could be considered because the trial would overlap, rather than reduce, the drug's effective market exclusivity period (Supplementary Discussion); and a post-marketing surveillance program might allow newly drug-treated individuals to enter the program on a rolling basis, replacing any who withdraw. The effects of these assumptions (Table 3) are collectively to reduce the number of individuals required to demonstrate efficacy of a drug with hazard ratio of 0.5 from 813 to 37. At the same time, the number of individuals available for a trial might increase, because: an approved drug should have broader geographic reach than a pre-approval trial; a treatment might improve awareness and diagnosis of the disease; and a treatment might stimulate more individuals to pursue predictive genetic testing. For instance, of people at 50/50 risk for a PRNP mutation, currently only 23% pursue predictive testing, compared to 60% (2.6X higher) for BRCA1 or BRCA2 mutations 37 , which are considered medically actionable 38 . Thus, a post-marketing study could be adequately powered with available numbers of individuals for a hazard ratio of 0.5 (Table 3) and, over a range of assumptions, would bring power requirements into closer alignment with the number of available individuals ( Figure S6 and Supplementary Discussion).

scenario N required explanation
Randomized pre-approval -5-year follow-up 813 See Table 2 Post-marketing with historical controls -5-year follow-up Increased power because the withdrawal rate is set to zero, simulating a scenario where individuals who go on drug can continuously enter the surveillance program, and the cohort being monitored can maintain its size over time. While we conducted tests to ensure that our power simulation was not itself biased (Supplementary Discussion), a post-marketing study could still be biased in real life, if the historical controls used do not accurately estimate the true hazard rates facing the trial participants 39 . There are no environmental, demographic, or non-PRNP genetic factors known to affect prion disease risk or age of onset, although these might nonetheless exist 40, 41 . Perhaps of greater concern is that most of our historical controls were collected retrospectivelyindividuals are only ascertained if they become sick -and may overestimate the hazard rates for individuals followed prospectively 32 . To assess this possible source of bias, we compared the survival of the limited number of individuals followed prospectively in our dataset (N=24 individuals, with a cumulative 145 person-years of follow-up), conditioned on their ages at first ascertainment, to those of individuals with no prospective follow-up. We did not observe a significant difference in hazard (P=0.59, Cox proportional hazards test) between these two groups, although this could be due to a lack of power (see Discussion).

Discussion
Both our power calculation and simulation indicate that direct demonstration of clinical benefit in a randomized pre-approval prevention trial would require enrolling a number of PRNP mutation carriers that is not currently realistic. For instance, for a drug that reduces annual risk of onset by half (hazard ratio of 0.5), estimated to correspond to a 7-year delay in median age of onset, 80% power was reached only with 813 individuals randomized for 5 years (Table 2). Currently, only N=221 presymptomatic individuals in the U.S. have positive genetic test results for PRNP mutations, and we estimate that only ~60 of these have high penetrance mutations and fall in an age range (≥40 years) where their hazard is sufficiently high that they would contribute appreciable power to a randomized trial with a clinical endpoint. Randomized prevention trials might just barely achieve 80% power under the most wildly optimistic assumptions of an extremely effective drug (hazard ratio of 0.1, reducing annual risk of onset by ten-fold), along with some increase in predictive testing rates and a very successful trial recruitment effort. FDA has cautioned, however, that rare disease trials should not be designed around the hope of a huge effect size 31 , and even if a drug were so profoundly effective, it is unlikely that a sponsor would have sufficient confidence in this a priori to invest in a trial that is underpowered for more moderate effect sizes.
At least three factors can explain why a randomized trial design following pre-symptomatic individuals to a clinical endpoint was deemed feasible for early-onset Alzheimer's disease, yet appears unviable for genetic prion disease. First, onset is less predictable in genetic prion disease. The standard deviation of age of onset ranges from ±10.0 to ±11.8 years for the three PRNP mutations we examined (Table 1), whereas estimates of the standard deviation of age of onset for PSEN1 E280A Alzheimer's disease range from ±6.4 to ±8.6 years 42,43 . In addition, an individual's age of onset in genetic Alzheimer's disease is reported to be correlated with parental age of onset 43 , and this property has been used to attempt to enrich for high-hazard individuals in trials 44 , whereas we have found no evidence that parent and child age of onset are correlated in genetic prion disease (Table S7 and ref. 32 ). Second, genetic prion disease is rarer. The PSEN1 preventive trial recruited from a single pedigree of ~5,000 individuals 45 from which 1,065 living individuals with the mutation have been enrolled in a registry 46 . There is no known genetic prion disease family this large. Third, genetic prion disease offers more limited financial incentives for a pharmaceutical sponsor. The cost of the PSEN1 preventive trial has been estimated at $96 million 44,47 , and while this price may be tenable for sponsors in view of potential for an expanded Alzheimer's indication, no similar potential exists for prion disease. Indeed, even in Alzheimer's disease, larger or longer primary prevention trials are likely to prove challenging for the private sector and may require public sector investment 48 .
Preclinical proof-of-concept studies in mice have shown that some antiprion agents effective at delaying prion disease on a prophylactic basis become ineffective if given close to the time of clinical onset 16,19 , suggesting that trials in symptomatic patients could fail to show a benefit that would have been realizable in preventive treatment. Yet our results here indicate that it would be difficult or impossible to design a well-powered randomized preventive trial with a clinical endpoint in genetic prion disease. Together, these observations argue for the characterization of biomarkers suitable as endpoints in presymptomatic genetic prion disease, and for their evaluation by regulatory agencies as surrogate trial endpoints. Accompanying manuscripts describe one possible route to Accelerated Approval using a surrogate biomarker endpoint 49,50 .
If Accelerated Approval could be achieved, then a post-marketing study would be required to confirm clinical benefit. We considered a model in which drug-treated individuals are enrolled in a surveillance program and their survival is compared to that of historical controls. We estimate that, compared to randomized pre-approval studies, such a program could reduce the number of individuals required for 80% power at the P=0.05 threshold by 3-to 20-fold. Meanwhile, conditional approval of a first prion disease drug may alter key parameters such as diagnosis, recruitment, and genetic testing rates, the last of which alone could increase participant availability by more than 2-fold. Thus, while power for any trial depends upon how effective the drug is, there exists a range of assumptions under which a post-marketing study could be adequately powered. There may be various formats through which the Accelerated Approval requirement of a post-marketing study to confirm clinical benefit could be met.
Under some assumptions, a post-marketing study might last a decade or longer and would benefit from following all mutation carriers taking the drug. With creative and careful planning, we propose that these goals could be achieved. In one model, a post-marketing study might take the form of a surveillance program, in which treated patients are followed long-term, perhaps in collaboration with existing prion specialist clinics and surveillance centers worldwide. In such a model, drug costs would be reimbursed by payors, in contrast to a more traditional sponsor-funded pivotal trial. While this model would be a departure from the more conventional design of most post-marketing studies required for recent Accelerated Approval drugs 2 , precedents exist for regulatory innovation in this area. For example, FDA's Risk Evaluation and Mitigation Strategies (REMS) program for drugs with serious safety concerns entails indefinite post-market enrollment and monitoring of treated patients 51 , and post-approval study requirements for medical devices often include registries or surveillance efforts and are not always industry-funded 52,53 .
Our study has several limitations. First, true age of onset distributions can only be obtained prospectively 54 , whereas our data are largely retrospective. We have included asymptomatic individuals with pathogenic PRNP variants where possible, but our ascertainment of them is certainly incomplete due to limited uptake of predictive testing 23 . This bias may tend to make our estimates of age of onset overly pessimistic 32, 55 . To the extent that true age of onset is older, or total lifetime risk lower, than our data suggest, randomized preventive trials with a clinical endpoint would require even greater numbers of individuals, and thus further increase our caution around this study design. Second, although our dataset is, to our knowledge, the largest ever reported for genetic prion disease age of onset, our statistical power to detect genetic modifiers, which might aid in age of onset prediction, is still limited. Third, although we have attempted to select a reasonable set of assumptions for modeling clinical trials, we have by no means exhaustively sampled the set of possible trial designs and parameters. Fourth, powering a post-marketing study will require a good historical control dataset to compare to, and our dataset, which was collected mostly retrospectively, may or may not be adequate. We found no evidence that our dataset overestimates the hazards facing prospectively followed individuals, but this could be due to a lack of power in our analysis. Fifth, the ascertainment of genetic prion disease by prion surveillance centers may be biased towards rapidly progressive phenotypes, meaning that the prevalence of more slowly progressive forms might be underestimated.
Our findings highlight two priorities for the prion field. First, the discovery and characterization of biomarkers capable of serving as trial endpoints may be essential to enable near-term presymptomatic trials in genetic prion disease. Second, a post-approval surveillance mechanism for age of onset merits consideration as one option for confirmation of clinical benefit in the context of Accelerated Approval. The ability to access therapies that can prevent or delay prion disease, yet which are likely to be less effective or ineffective after symptom onset, could be greatly enhanced by success in these areas.

Methods
Literature annotation. We considered 69 reportedly pathogenic PRNP variants (Table S1) and reviewed primary literature to determine which had evidence of at least one family with at least three affected individuals in a pattern consistent with Mendelian segregation, or had a documented case with a de novo mutation. We identified 27 such variants, deemed likely high penetrance variants. The remainder were seen in isolated patient(s) with a negative or unknown family history, and/or have population allele frequencies inconsistent with high penetrance 6 . These variants will include both benign and low-risk variants. It is possible that some genuinely high penetrance variants may also lack literature evidence for high penetrance due to missing family history information or an unavailability of family member DNA to confirm de novo status, but this issue will only affect variants with very low case counts and thus will have minimal impact on the results reported here.
Data collection. Age of onset data were gathered from nine study centers: the UK National Prion Clinic, the German Reference Center for TSEs, the Memory and Aging Center at University of California San Francisco, the Australian National CJD Registry, the reference center for CJD at University of Bologna, the DOXIFF study at the Mario Negri Institute, the Japanese national prion surveillance network, the French national reference center for CJD, and the Spanish National Center for Epidemiology. The data include both previously reported and newly identified families and individuals. Data were collected through clinical visits, reports to prion surveillance centers, and family histories, as previously described 32,56-61 . Age of onset was based on the earliest date of symptoms, determined by the patient or witnesses, that subsequently developed into prion disease. Data on the number of positive predictive genetic tests for PRNP mutations was provided by the National Prion Disease Pathology Surveillance Center for this study.
Life tables and hazard curves. We tabulated, for each PRNP mutation and for each age from 1-100, the number of individuals alive at the beginning of the interval (lives; l), becoming sick or dying within the interval (deaths; d), or being censored -alive and well at last followup or dead of a different cause -within the interval (withdrawals; w). The raw hazard (q) was computed as onsets divided by the mean number of people observed over the interval: q = d/(l -w/2), and a smoothed hazard (q_smooth) was computed by passing a Gaussian filter (sd=3 years, maximum width=15 years) over the raw hazard. The proportion surviving for each interval (p) was 100% for the first year and was computed as (1-q) times the proportion surviving in the previous interval for every year thereafter. To compute the 95% confidence intervals on the smoothed hazard, we sampled each mutation's data, with replacement, 1000 times, generated life tables for iteration, and then chose the 2.5 th and 97.5 th percentile of the hazards in the bootstrapped distributions at each age.

Assumptions.
To determine a reasonable assumption for withdrawal rate, we performed Google Scholar searches for preventive trials in neurology (N=2) or cardiology (N=6). The annual withdrawal rate was computed as w = 1 -exp(log(A)/t)), where A is the proportion of patients completing the trial at time t. Results are summarized in Table S3.  6% for Table 2), and hazard for the drug group was the baseline hazard times the hazard ratio. The cumulative event rate in each group was computed as C = (h/(h+w)) * (1-exp(-(h + w)*t))), where h = hazard, w = withdrawal rate, and t = years of followup. The overall cumulative event rate Ctot was the average of the cumulative event rates for the two groups, weighted by proportion treated (in this case, 50/50). The number of randomized individuals required for d events to be observed was calculated as d /Ctot. To account for ignoring the first g years of data, we reasoned that the cumulative rate of events usable in the final dataset would be Cusable = (h/(h+w)) * (1-exp(-(h + w)*t))) -(h/(h+w)) * (1-exp(-(h + w)*g))), which simplifies to Cusable = (h/(h+w)) * (exp(-(h + w)*g)exp(-(h + w)*t)) Simulations   Figure S1.

Estimation of number of individuals available for trials
It is possible to estimate the true number of high penetrance PRNP mutation carriers based on disease prevalence. Using data from recent case series, 1,176 prion disease cases harbored a PRNP variant classified here as highly penetrant, out of 10,460 sequenced cases or 16,025 total cases 6 . Thus, 7 -11% of prion disease cases have a high penetrance PRNP variant. Prion disease is responsible for ~1 in 5,000 deaths 6 , suggesting that ~1 in 45,000 to 71,000 deaths are due to a high penetrance PRNP variant. The carrier rate among the living population will be somewhat lower because these variants reduce life expectancy, but it is reasonable to suppose that ~1 in 100,000 people harbors a high penetrance PRNP variant. This is in line with recent population control data, where out of 138,632 individuals in the gnomAD database as of December 2017 (http://gnomad.broadinstitute.org/) 63 , there is one individual with the E200K mutation and no others with any variant classified here as high penetrance. Similarly, out of ~531,575 individuals genotyped by 23andMe, between 1 and 5 harbored one of four well-known high penetrance variants (P102L, A117V, D178N, and E200K) and between 1 and 5 harbored one of an additional set of variants which includes three classified here as high penetrance (P105L, T183A, and F198S). If the true carrier rate is 1 in 100,000, then there may exist 3,000 people in the United States with high penetrance PRNP variants. However, this figure greatly overestimates the number of people available for trials, as most of these individuals have not undergone predictive testing. Indeed, many are likely not even aware that they are at risk, perhaps because a family history is absent or a family member was not diagnosed correctly.
The National Prion Disease Pathology Surveillance Center in Cleveland, Ohio, as the sole provider of PRNP gene testing in the U.S., has exhaustive ascertainment of individuals who have chosen predictive testing for genetic prion disease in this country. In the period from 1996 through January 2017, it provided N=221 positive predictive test results, for any PRNP variant, to individuals who are not known to have developed disease as of 2017. Privacy concerns prevent publication of a breakdown of this number by age and specific PRNP mutation, but estimates can be made based on other cohorts. Among U.S. symptomatic prion disease cases with a rare PRNP variant, 75% (271/362) of individuals had a mutation classified here as high penetrance (Table S1), and in the reported U.K. predictive testing cohort 23 , 36% (37/104) of individuals who chose predictive testing were age 40 or older. Thus, a conservative estimate that there are only 221× 75%×36% = ~60 individuals alive in the United States today who meet the criteria we use in our power calculations.
The above estimate is conservative in that it reflects individuals who currently know their genetic status. In the U.K. predictive testing cohort, only 23% of individuals at 50/50 risk chose predictive testing 23 , similar to reported figures in Huntington's disease, another incurable neurological disease (see refs in 23 ). In contrast, 60% of individuals at risk for BRCA1 or BRCA2 mutations, for which preventive measures are available, chose predictive testing 37 . Thus, it is possible to imagine a 60%/23% = ~2.6X increase in the uptake of predictive testing if a preventive therapy for prion disease were available. Thus, a more generous estimate of the number of individuals age ≥40 available in the U.S. is 60×2.6 = 156. Such an estimate is probably more realistic when considering an approved prevention measure (as in postmarketing studies) than when considering an experimental drug entering randomized preapproval trials (see below).
Although we contemplated worldwide trials with multiple international sites, we did not have adequate data to estimate the number of genetically tested presymptomatic individuals worldwide. The NHS National Prion Clinic in the U.K. has seen 72 presymptomatic individuals with PRNP mutations since 1990, and the French surveillance center in Paris has delivered 18 positive PRNP predictive test results since 2004, but the other centers involved in this report did not have comprehensive data on predictive testing in their respective countries analogous to that available for the U.S. We also note that a large number of E200K mutation carriers are suspected to exist in Slovakia and Israel due to founder mutations, although fewer than 100 carriers appear to have been identified in each country to date 64,65 .
We also considered estimates based on the incidence of genetic prion disease. U.S. prion surveillance reported 271 individuals dying of prion disease with high penetrance mutations, suggesting that at least a comparable number of carriers in the U.S. are currently healthy and will have onset with a correct diagnosis within the next 15 years. The comparable figure including Europe, Australia, and Japan is 1,176. These last two figures are still lower than the true number of carriers in existence due to underdiagnosis, yet they overestimate the number of individuals actually reachable for trials because they ignore the question of how many individuals would choose predictive testing, and they include individuals who would be difficult to ascertain prospectively because they lack a known family history of prion disease, either due to de novo mutations, incorrect or incomplete information about previous family illnesses, or <100% penetrance.
For all of the above estimates, an important caveat is that the number of individuals successfully recruited, screened, and enrolled for a trial will be only a fraction of the number who meet the most basic enrollment criteria such as genetic status and age. Willingness, geography, and various exclusion criteria will dramatically lower the number actually enrolled.
Finally, it is worth noting that all of our calculations and assumptions are based upon the present moment, when there exists no drug for prion disease. It is likely that approval of a first prion disease drug would increase the number of patients available for future trials. A drug could improve diagnosis rates, as prion disease is not currently prioritized in the differential diagnosis of rapidly progressive dementia due to its being untreatable 66 . The U.S. observes an incidence of ~1 prion disease case per million population per year, but up to twice that incidence has been observed in countries with more intense surveillance systems 29 . Because many prion disease patients die undiagnosed, their relatives may never learn that they are at risk for a PRNP mutation. A drug might also increase the uptake of predictive genetic testing among those who do learn that they are at risk. The 23% uptake observed for prion disease 23 is consistent with other currently "medically inactionable" indications such as Huntington's disease 67 , while as noted in the main text, "actionable" indications such as BRCA1/2 mutations appear to have much higher uptake 37 . Finally, the existence of a drug may promote general awareness of the disease and improve the infrastructure for surveillance, registries, and patient ascertainment.

Simulation of power for randomized preventive trials with a clinical endpoint
Individuals were assigned one of the three PRNP mutations and a starting age distributed between 40 and 80, weighted by mutation prevalence and by the proportion of individuals surviving at each age. As above, we assigned half of individuals to drug and half to placebo, and assumed a w=15.2% annual withdrawal rate, a P=0.05 statistical threshold, and a 5-year trial duration with a 1-year "run-in" period. For each year of the trial, each individual withdraws with probability w, becomes sick with a probability corresponding to the hazard function for their particular PRNP mutation and age at the time, multiplied by the simulated hazard ratio if drug treated, or else continues on in the trial. At the end of each simulated trial, we analyzed the censored trial data to determine a P value. For non-stratified simulations, drug/placebo status was assigned without regard to mutation, and survival status was regressed on drug/placebo status alone using a log-rank test, with the overall P value as the readout. For stratified trials, drug/placebo status was assigned 50/50 within each mutation, and mutation was included as a covariate in a Cox proportional hazards regression, with the P value for the "drug" parameter as the readout.
We then compared this model to the power calculation results by taking the calculated required numbers of individuals for 80% power ( Table 2) and then running the simulation (500 iterations) to determine the power for this number of individuals. The results (Table S6) show overall good agreement between the power calculation and the simulation -for most scenarios tested, the power is indeed close to 80%, with or without stratification. Stratification actually reduces statistical power for the conditions with low N and low hazard ratios. Under such conditions, it is a common occurrence that there may be zero disease onsets either in one randomized group (usually the drug-treated group) or in one mutation, resulting in an infinite regression coefficient or beta in the Cox model. Thus, the regression never converges, and the simulated trial results turn out statistically non-significant.

Codon 129 effects on age of onset and disease duration
To determine whether codon 129 affects age of onset for the three most prevalent mutations considered here, we used a log-rank model based on codon 129 diplotype (phased genotype) where available (Table S7). In this model, only D178N showed clear evidence for genetic modification of age of onset and disease duration, with P values significant after multiple testing correction. To determine the nature of this genetic modification, we plotted survival curves by codon 129 diplotype and, because phase was unknown for many codon 129 heterozygous individuals, we also considered phaseless genotypes. In pairwise tests for D178N, M/M was not significantly different from M/V (nominal P = 0.14) nor from V/M (nominal P = 0.69), and in the phaseless survival curve, MV was overall similar to MM ( Figure S4D). These results suggest that the significant codon 129 effect on D178N age of onset is most likely driven primarily by a younger age of onset in V/V individuals compared to other diplotypes. Despite the strong statistical significance of this difference, the small number of D178N-129VV individuals means that codon 129 does not add any explanatory power for age of onset in the dataset as a whole. As noted in the main text, mutation alone explains limited variance in age of onset (adjusted R^2 = 0.15, P = 1.3e-33). Adding cis and trans codon 129 to this model decreases the variance explained (adjusted R^2 = 0.14, P = 3.6e-18).
We also investigated in further detail previously reported associations. For disease duration, D178N M/M and V/V were significantly more rapid than either heterozygous diplotype, consistent with previous reports. Although codon 129 diplotype did not have a significant effect on E200K disease duration overall (nominal P = 0.10), a phaseless genotypic model was suggestive (nominal P = 0.031), with MV heterozygotes appearing to have a slightly longer disease duration than MM homozygotes, a direction of effect consistent with previous reports 5,32 . Whereas P102L age of onset was reported to be higher for M/V than M/M individuals 36 , here we find no evidence for this and, ignoring phase, the non-significant trend is towards younger onset in MV than MM individuals (nominal P = 0.056).

Potential age of onset confounders
Because our data were gathered from a variety of study centers using a variety of methodologies, we asked whether any confounders might affect age of onset (Table S7). There was no difference in age of onset between directly and indirectly ascertained individuals (P = 0.78). Age of onset was correlated with year of birth after controlling for mutation (P < 1e-48), which is a previously reported artifact caused by our relatively limited ability to ascertain individuals whose onset has not yet arrived (though we ascertain some of them through predictive testing) or whose onset occurred before genetic diagnosis of prion disease was possible (though we ascertain some of them through family histories) 32 . This correlation does not affect estimation of overall age of onset distributions. Age of onset appeared to differ slightly among the nine contributing study centers after controlling for mutation, although it was not significant after multiple testing correction (nominal P = 0.012, Bonferroni P = 0.26, two-way ANOVA), and it only marginally increased variance explained (adjusted R^2 = 0.16) compared to mutation alone (adjusted R^2 = 0.15, see above).
Year of onset showed evidence of positive correlation with age of onset after controlling for study center and mutation (nominal P = 0.00032, Bonferroni P = 0.008, linear regression), although the effect size was small (+0.12 years of age per calendar year, or in other words, cases in 2010 have on average an age of onset 1.2 years older than cases in 2000) and, again, the impact on variance explained was minimal (adjusted R^2 = 0.18). This slight positive correlation might be due to improved ascertainment of older-onset cases as prion surveillance strengthens over time.

Justification for trial duration assumptions
In the main text, we argued that a longer trial duration could be considered for a post-marketing study because it would run concurrently with, rather than reducing, the drug's effective market exclusivity period (the period before generic equivalents can be approved). In the U.S., new drugs may be protected by patent exclusivity granted by the Patent and Trademark Office and/or by market exclusivity measures granted by FDA; these exclusivity periods are not additive. Patents last 20 years beginning from their filing, which is usually during the preclinical development phase. The 1984 Hatch-Waxman Act allows sponsors to recover up to 5 years of additional exclusivity, not to exceed a total of 14 years of market exclusivity, to make up for time the drug spends in FDA review 68 . FDA can offer varying periods of market exclusivity depending upon the indication and treatment modality, including 12 years for new biologics 69 and 7 years for rare disease drugs granted Orphan Drug designation 70 . In practice, new drugs receive on average about 12 years of effective market exclusivity 71,72 . The vast majority of pivotal trials supporting new drug approvals last less than one year 73 . While there are rare examples of 5year trials 47 , a 10-or 15-year prevention trial would exhaust most or all of a drug's effective market exclusivity period. In contrast, as noted in the Discussion, there do exist precedents for very long-term surveillance of patients receiving a drug after approval.

Historical control trial simulation
As for the simulation of randomized trials, individuals were assigned one of the three PRNP mutations and a starting age distributed between 40 and 80, weighted by mutation prevalence and by the proportion of individuals surviving at each age. Again, we assumed a w=15.2% annual withdrawal rate (Table S5), a P=0.05 statistical threshold, and a 1-year "run-in" period where disease events are ignored. Distinct from the randomized trial simulation, here all simulated individuals are treated with the drug. For each year of the trial, each individual withdraws with probability w, becomes sick with a probability corresponding to the hazard function for their particular PRNP mutation and age at the time, multiplied by the simulated hazard ratio, or else continues on in the trial. At the end of each simulated trial, the censored trial data on treated individuals are compared to our original dataset as historical controls (Supplementary Life Tables). To determine a P value we used a Cox proportional hazards counting model accounting for different left-truncation times 74 : for untreated individuals in the original dataset, we assumed age 0 as a start time, while for treated individuals, we assumed left truncation at the age at trial enrollment, plus one year to account for the "run-in" year.
While we cannot currently rule out the possibility that our dataset is biased relative to the true hazards facing mutation carriers in real life (see main text Discussion), we sought to confirm that our simulation method is not itself biased. We reasoned that if our simulation was unbiased, then for a drug with hazard ratio equal to 1 (a completely ineffective drug), even long trials with large numbers of individuals should have power equal to alpha, by the definition that alpha is the false positive rate when the null hypothesis (no efficacy) is true. We therefore ran 1000 iterations of a simulation with a hazard ratio of 1 and 1000 individuals followed for 20 years. We observed a significant result at P < 0.05 in only 5.5% of iterations, consistent with the expected 5%.
In contrast to the result for randomized trials (see discussion above and Table S6), we found that stratification by mutation in the analysis of historical control trial simulation did just slightly increase statistical power. For example, with N=156 individuals followed for 15 years with a hazard ratio of 0.5, power was 90.6% (906/1000 iterations) without stratification and 94.1% (941/1000 iterations) with stratification. This difference from the randomized trial simulation may be a property of the Cox counting model, combined with the fact that our historical comparison dataset has N=1,000 individuals, and we considered follow-up periods of up to 15 years, meaning that the dataset was large enough for the small explanatory power of different PRNP mutations to matter. Nevertheless, for consistency with the methods used for the randomized trial simulations, we chose not to stratify in the simulations used for Table 3 and Figure S6.
We performed power calculations for post-marketing studies using historical controls under a range of assumptions in addition to those explored in Table 3 in the main text. In one set of experiments, we considered the effects of varying the length of the follow-up period. For a hazard ratio of 0.5, 80% power could be achieved within 9 years for N=156 participants, but is never achieved for N=60 participants ( Figure S6A). This is because statistical power eventually plateaus for lack of participants: our assumption of a 15.2% withdrawal rate compounded annually means that after 10 years, only 19% of the original participants remain in the trial. If the set of drug recipients followed in a post-marketing study were fixed shortly after approval, then this is a realistic concern. If, on the other hand, study design allows new individuals who are prescribed the drug to be added to the monitored cohort continually, the number of individuals in the trial could stay constant or even grow. To simulate this possibility, we also considered a zero withdrawal rate scenario. Under this assumption, even with N=60 individuals, 80% power is achieved in 10 years ( Figure S6A).
In another set of experiments, we compared the power for post-marketing studies with historical controls, with or without modeling withdrawal, in comparison to pre-approval randomized trials, for a range of hazard ratios ( Figure S6B). For the same hazard ratio and level of statistical power, post-marketing trials generally required only about one fifth as many individuals, and if withdrawal is set to zero, simulating continuous enrollment, only one twentieth as many, as preapproval randomized trials.
Certainly, a post-marketing study is not a panacea, and under certain assumptions even this trial design is not well-powered: for instance, for a drug of marginal efficacy (hazard ratio 0.9, delaying onset by ~1 year) even a 15-year trial with no withdrawal could not achieve 80% power with 1,000 participants. But, under a range of moderate assumptions, a post-marketing study is more feasible than randomized pre-approval trials with a clinical endpoint.     Table 2). A) Best case scenario: overall average hazard is 4.8% (the higher figure including the less common mutations shown in Table S2 and Figure  S1), the withdrawal rate is 6.9% per year (the lowest rate in any of the trials we reviewed, see