Drug Development in Rare Diseases: Need for Innovation in - - PowerPoint PPT Presentation

1

Drug Development in Rare Diseases: Need for Innovation in Statistical Thinking

Kannan Natarajan Nov 6, 2019

2

Seeking New Treatment in Rare Disease

Source: PhRMA 2013 report on Rare Diseases

3

Current Landscape

https://www.pfizer.com/science/rare-diseases

4

Challenges with Traditional Development in Rare Diseases

• Few patients available to participate

– Multi-center, multi-country trials

• Unmet medical need
• Phenotypic diversity and genetic subsets

– heterogeneity at presentation / late diagnosis – highly variable disease course

• Lack of well defined and validated endpoints, outcome measures/tools,

and biomarkers

• Natural histories are often not well understood/characterized

5

Changing Regulatory Perspectives

• 21st Century Cures Act,

December 2016

– includes provisions that will improve the development, for rare disease patients – further expansion of the Patient-Focused Drug Development

• New FDA guidance

– use of natural history data – human Gene Therapy for Rare Diseases

6

Need for Innovative Study Designs

• Design considerations

– endpoint: asking the right question – Replace or augment control arm with available information in standard of care – Extrapolation to other demographic subgroups – modeling disease from natural history data – use of real world data (RWD) Other considerations

• collaboration and data

sharing

• registry development

7

Innovative Design in Drug Development

• An innovative trial design uses all available evidence for better and

efficient drug development without undermining the validity and integrity. The goal is to provide fast access of good drugs to patients. Validity  providing correct statistical inference:  providing convincing results to a broader scientific community  minimizing statistical bias Integrity  Pre-planning appropriately  maintaining confidentiality of data  minimizing operational bias

8

Use of Historical Control Data

9

Use of Historical Control

• More than 220 000 registered trials in the electronic database

– typically more than one clinical study investigating the same treatment

• Using available information in design and analysis may lead to

– increase efficiency: fewer patients – ethical – decreased costs and trial duration

• It is useful particularly when

– information is sparse (e.g. rare disease) – new information is difficult to obtain

• Recommendations and considerations are provided in ICH E10

10

Historical Data Augmentation Design

• Use of available information in the design

– Use information for control worth n* patients and allocate n- n* patients – saves sample size

• Choice of relevant historical data?

– Requires judgment about similarity of historical and current setting

• Inclusion/exclusion criteria, endpoint, time-

trends etc.

– Requires interaction with non-statistician

• If information available after trial starts

– Indirect comparison: evidence synthesis

S c r e e n

E C HC

SS=n SS=n – n* pESS=n* End of Study

11

Choice of “Relevant” Historical Data

• Proper choice of historical data: justification
• f “similarity”

– prior to start of trial – choice must be “science” based not “result” based – avoiding publication bias: often requires KOL and independent groups – inter-disciplinary collaboration – data gathering can be time-consuming

• Pocock’s (1976) “six criteria”

12

Key Statistical Considerations

• Both frequentist and Bayesian methods available

– Few frequentist methods include Test and Pool, Propensity Score

• Bayesian methods provide a natural way to incorporate historical data in

the form of prior

– handles the between trial heterogeneity

• Various available methods:

– Meta-analytic approaches are well established to handle both (Neuenschwander et. al 2010)

• Meta-analytic framework provides a flexible framework to incorporate

relevant data from different sources

13

Checklist for Practical Implementation

• Statistically principled approach

– Incorporating between trial heterogeneity – Quantify «borrowing»: how much to borrow – What if historical data and actual data are in conflict?: requires robust statistical approach (Schmidli et al. 2014, Neuenschwander, Roychoudhury and Schmidli 2016)

• Evaluation of frequentist operating characteristics (Type-I error, Power)

–

Wide number of scenarios

• Protocol and manuscript writing
• Communication with key customers

14

Meta-analytic Approaches

Derivation of informative prior using historical data: two-step approach Meta-Analytic Predictive(MAP) Meta-Analytic Combined (MAC) Combined analysis of historical data and current data: one analysis MAP and MAC are equivalent: “exchangeability” is the key assumption

• Methodology is general enough to be extended for different endpoints

(continuous, binary, count data, time to event)

15

Robustness to Handle Prior-data Conflict

Prior-data Conflict Scenario

De Groot always carried an ε of probability for surprises in his pocket! Robustness and more rapid adaptation to prior-data conflicts by adding extra weakly-informative mixture component

16

Current Regulatory Consideration

• Recently released regulatory guidelines encourage the efficient use of

historical data

– Primary concerns theoretical strong control of Type-I error, bias etc. – Need upfront discussion with regulatory authorities

• Few exceptions: regulatory approval using historical control

– Brineura for Batten Disease – APTIOM as monotherapy for Seizures – Venetoclax in Relapsed / Refractory Chronic Lymphocytic Leukemia (CLL) – Eteplirsen in Duchenne Muscular Dystrophy (DMD) – Suvodirsen in DMD: submitted for FDA CID program

17

Example: Progressive Supranuclear Palsy (PSP)

• Progressive supranuclear palsy (PSP)

is a degenerative neurologic disease due to damage to nerve cells in the brain

• 20,000 PSP patients have been

diagnosed with the disease (6·5 cases per 100 000 individuals)

• No effective drug halting the

progression of the disease

18

Example: Phase II Trial in PSP

• Disease

– PSP

• Experimental treatment

– Monoclonal antibody (E)

• Endpoint

– PSPRS (A clinical rating scale) change from

baseline assessed at week 52

• Traditional clinical trial design

– New treatment (n=80) vs. Placebo (n=80) – Z test

Can historical placebo information be used?

Golbe and Ohman-Strickland 2007

19

Example: Use of Historical Placebo Data

• Double-blind, randomized,

placebo-controlled study for E

• Primary endpoint: Change from baseline in

PSPRS at 52 weeks

– 4 points from placebo clinically meaningful

• Planned sample size 120

– 2:1 in favor of E – Z test: 73% power for δ= 4

• 2 historical trial data for placebo (n=144))

– Tideglusib vs. placebo (Tolosa et. al. 2014) – Davunetide vs. placebo (Boxer et. al. 2014) Study N Y se Tolosa14 21 10.4 6.5 Boxer14 123 10.9 11.0

20

Example: Informative Prior for Placebo Arm

• Historical data for placebo is homogeneous

in two studies

– Sample size varies: poses uncertainty

• MAP prior reflects this uncertainties

– apriori placebo effect varies 7.8-13.3 – prior worth 52 subject information for placebo – non-informative prior for E

• New trial is successful if

– P( δ < 0 | data) > 97.5%

21

Example: Robustification of MAP Prior

• A mixture of MAP and weakly informative prior becomes a heavy-tailed

version of the historical prior

– Mixture prior- 75% informative, 25%non-informative

MAP prior (100%-0%) Robust MAP prior (75%-25%)

22

Example: Robust Prior Handles Prior Data Conflict

Note: Weights are fix apriori but posterior weights get updated using standard Bayesian calculus (Schmidli et. al 2014) Scenario: No Conflict Scenario: Conflict

Weights

• aprior

informative 75% weak 25%

• postrior

informative 90% weak 10% Weights

• aprior

informative 75% weak 25%

• postrior

informative 1% weak 99%

23

Example: Operating Characteristics

• Robust prior provides a nice balance between

Type-I error and power

− Type-I error: well controlled when prior and data are aligned − Type-I error: max 8% under prior-data conflict − Power= 87% for δ= 4: considerable gain over traditional frequentist design

• Type-I error inflation is much higher with

informative prior only under prior-data conflict

24

Extrapolation to Demographic subgroups

25

Extrapolation of Efficacy from Adult to Pediatric Population

• Drug development for pediatric rare disease faces substantial hurdles,

including economic, logistical, technical, and ethical barriers

• An efficient design for rare pediatric population may “extrapolate”

– from adults to pediatric patients, between pediatric subpopulations

• Extrapolation can be considered as an extension of “borrowing historical

control data”

– extrapolation from adult population to pediatric refers to borrowing “treatment effect” information

• Assumes that there is no need for formal proof of efficacy in the pediatric

population

– no substantial difference in proof of mechanism between adult and children (supporting PK/PD information)

26

Regulatory Perspectives

• EMA (2018) : Reflection paper on extrapolation of efficacy and safety in

pediatric medicine development ... using Bayesian methods to either summarize the prior information for the extrapolation concept, or to explicitly borrow information (from adult trials, from control groups, from other pediatric clinical trials)

• FDA (2016): Leveraging existing clinical data for extrapolation to

pediatric uses of medical devices While Bayesian methods are described in this document, non-Bayesian methods can also be used for borrowing strength

27

Extension of Meta-analytic Framework

• Meta-Analytic framework: a powerful tool for

extrapolation

– flexible structure of borrowing from different cohorts (adults, adolescents, and younger children)

• However validation of extrapolation concept is

key

– use of predictive check to ensure data or model adequacy for extrapolation

• Predictive Evidence Framework: provides a

measure of adequacy of information for regulatory purposes (Neuenschwander, Roychoudhury, and Branson 2017)

Gamalo et. al. 2019

28

Example: Pediatric Trial on Guillain-Barre Syndrome

• Guillain-Barre syndrome (GBS) is a rare disease
• Affecting about 6,000 to 9,100 people in the U.S. each year

– annual incidence rate of GBS was 1.51 per 100,000 children – most common paralytic illness of children

• Objective: Efficacy assessment of Plasmapheresis in children
• Plasmapheresis is a complex procedure to do on children

– it is difficult to perform well powered study for pediatric population – can one use the available adult population information?

• Endpoint: time to recover ambulation

– defined as the ability to walk at least 5 m without assistance (grade 2 or lower)

29

Example: Adult and Pediatric Data of Plasmapheresis (PE)

• Two large randomized trials PE

vs placebo (A1 and A2)

– substantial decrease in the median time to unassisted walking – Highly significant HR

• Four small observational

pediatric trials (C1-C4) are available

– evidence is promising but sparse!

N HR 95% CI

C4 15 1.52 (0.54, 4.29) C3 19 0.55 (0.23, 1.34) C2 24 0.40 (0.16, 1.03) C1 23 0.40 (0.17, 0.94) A1 245 0.62 (0.46, 0.84) A2 220 0.63 (0.47, 0.84)

30

Example: Extrapolation of Adult Data to Pediatrics

• Cumulative results for the

predicted HR for a pediatric confirmatory trial

– starting with trial A1 then adding

• ther trials: trials A1+A2, trials

A1+A2+C1, etc. – using meta-analytic framework

• Strong adult data combined with

sparse children data can justify an efficacy claim for children

31

Use of Real World Evidence

32

Use of Real World Evidence (RWE): Widening the Scope

Real World Data (RWD) Real World Evidence (RWE) potential benefits or risks, of a medical product derived from analysis of real world data (RWD)

33

Real World Data (RWD): Heterogeneous Data Sources

• Rich data routinely collected from various heterogeneous sources

– a great source of information to enrich rare disease drug development

• Potential use of real world data includes

– an external control or internal concurrent control – extrapolation of treatment effect

• This will require adjustments to the methodology discussed earlier with

regard to bias and heterogeneity

– having access to relevant predictors that explain anticipated biases – meta-regression (partial exchangeability) – use of propensity score: matching, inverse probability treatment weight – often, a large number of variables: requires machine learning techniques

34

Design of Rare Disease Trial with RWE

• RWE can improve efficiency of drug development in rare disease

– Randomized delayed start (RDS) – Randomized enrichment with RWE control (RWE-RE) – two stage designs: combining two stages requires intra-patient comparison – received support from regulatory agencies in rare disease clinical trials

• Use of RWE in regulatory aspects requires further caution

– Heterogeneous nature of data: difference in endpoint data collection – Missing data

• Need upfront discussion with health authorities

35

Natural History Data: Understanding the Disease Mechanism

• Disease progression can be modelled based
• n natural history data
• This can be used to evaluate, inform, and
• ptimize clinical trial design

– projecting required sample sizes – identifying relevant patient populations – estimating the magnitude of treatment effect – defining the required duration of follow up – Help to identify relevant “surrogate” of disease progression – helps trial simulation

• DIAN: a nonlinear function captures disease

progression (Wang G et al 2018)

Quintana M, CTMC Webinar 2019

36

Use of Master Protocols

37

Master Protocol Design in Clinical Trial

• PDUFA VI and 21st Century Cures Act Commitment

– Complex Innovative Design project – FDA pilot program: Sponsors gain increased interaction with agency on design

• “Disease characteristic-based” designs (e.g. Biomarker)

predictive biomarkers use of new therapy

– Interaction designs, Enrichment designs, Adaptive enrichment designs, Marker strategy designs – Master protocol

• Personalized medicine
• Increased efficiency in drug development when target-drug combinations exist

38

Basket, Umbrella, and Platform Trial

• Multiple diseases, multiple patient subgroups (biomarker-defined),

and/or multiple therapies studied under one, over-arching protocol

• Different types of Master protocols

– Basket trial – Umbrella trial – Platform trial

• Current example: “most” Oncology

– Have potential in other therapeutic area

• The terminologies are not straight forward

– Early literature use “basket” , “umbrella”, and “platform” trial inconsistently

39

Master Protocol: Basket, Umbrella and Platform

https://sms-oncology.com/news/blog/the-changing-landscape-of-oncology-clinical-trials-aacr-2017/

40

Master Protocols Terminology

Master protocols

Master Protocols A single, over-arching protocol designed to study multiple patient populations, multiple diseases, multiple therapies, or all of these

Basket Trial Umbrella Trial Platform Trial

Objective To study the activity of a single therapy in the context of multiple diseases or disease subtypes. Note that a single therapy could be, for example, a combination of study drugs or other therapeutic interventions. Objective To study the activity of multiple therapies in the context of a single disease that could include multiple disease subtypes and therapies specific to the disease subtype. Objective To study multiple targeted therapies in the context of a single disease (but allowing also for multiple disease subtypes within a single disease) in a perpetual manner, where therapies may be added or dropped from the “platform” on the basis of a pre-specified decision algorithm.

41

Basket Trial: Key Features

• Typically used in early phase

– Single-arm sub-studies

• Preliminary target-treatment hypotheses

– Often uses short term efficacy endpoint

• Target: identify large, unambiguous signals of activity based on

molecular features (rather than tumor type)

–

“Success” within a sub-study may lead to larger confirmatory study

• Usually 20-30 patients per “basket” or molecular sub-study
• Model-based methods for pooled analysis of multiple disease is useful

– Bayesian Hierarchical Model or Robust Hierarchical Model

42

Case Study 4: Basket Trial Design in Sanfilippo Syndrome

Child with disease sub-type Arm 1 (A) Arm 2 (B) Arm 3 (C) Arm 4 (D) Interim Analysis Perform Final Analysis Primary Endpoint: Disease Control Rate (DCR) A disease sub-type:

• STOP for futility
• Continue to Final

Analysis

43

Umbrella Trial and Platform Trials: Key Features

• Mid-to-late phase sub-studies

– Typically randomized design for sub-studies – Concurrent or common control – Often includes futility or efficacy interim analysis

• Better understood target-treatment hypotheses
• Objective remains identification of large effects

– Careful consideration of trial size for different sub-study

• The final analysis can be traditional or novel depending on the design

and objective

• Predefined rule for early futility or “graduation”
• Adaptive on-trial changes to reflect changing treatment landscape

44

Example: ISPY-2

https://www.onclive.com/publications/oncology-live/2014/march-2014/i-spy-2-designer-describes-programs-many-innovations

45

Example: IMI-EPAD and DIAN

• IMI-EPAD and DIAN are platform trials in Alzheimer Disease
• IMI-EPAD

– Patients enter a longitudinal cohort study (LCS) and then can be directed into the PoC platform trial

• DIAN

– Platform trial for patients with dominantly inherited Alzheimer – Strong historical data used to build disease progression model

• Both trials leverage the use of placebo controlled patients borrowed

across the treatment cohort

– Decreases the number of patients in placebo

Source: Satlin et. al. 2016

46

Master Protocol Implementation

• Prospective planning is essential
• Need to explore and present multiple scenarios to team
• Operational aspects

– How to incorporate change in practice?: new SOC – Protocol amendments – Multiple IA, unblinding, DMC, and data integrity

• Enrollment rate: has great impact
• Extensive simulation of trial performance

– Type-I error, power and graduation probabilities

47

Other Designs: snSMARTs

• Patients are guaranteed to be on at least
• ne drug.
• Determine best and second best drugs in

Stage 1 (A and B) based on response criteria

• A vs B Contrast from Stage 1 and from C

non-responders in Stage 2 are weighted and pooled

• Frequentist and Bayesian framework for

modeling is available for snSMARTs design (Tamura 2016, Wei et. al. 2018)

Wei et al 2018

48

Estimands in Rare Disease Trials

49

Estimand in Rare Disease: Asking the Right Question

• Recently issued draft ICH E9 (R1) addendum asks clinical trials to

clearly describe the treatment effect to be estimated or estimand

• Estimand framework handles the intercurrent events that may

complicate estimation of treatment effects

– addendum outlines five key strategies for intercurrent events – treatment strategy and hypothetical are two most common strategies

• Estimand framework is important in the rare disease setting as well

– important question of interests are clearly stated in protocol – sample size is adequate to address this primary estimand of interst with proper scientific rigor

50

Change from Baseline or Longitudinal Analysis?

• Assessing disease progression in longitudinal fashion provides higher

power than traditional change from baseline

– they address different questions (“estimand”) – change from baseline: ““what is the effect at a given time” – longitudinal analysis: “how the treatment effect develops” – important to evaluate the need

• Four attributes need to clearly stated in protocol
• Example:

– The primary estimand will be the difference in mean change from baseline at Week xx in the <variable of interest> between drug A and placebo, in all randomized participants, regardless of <occurrence of intercurrent event> during the first xx weeks.”

51

Use of Composite Strategy

• Composite strategy in rare diseases increase power and/or to sufficiently

capture complexity

– TTR Amyloidosis: all cause mortality and frequency of CV related hospitalizations – Fabry disease: cardiac events, renal events, or related death

• Does not suffer from the strong assumptions behind treatment policy and

hypothetical estimand

– it includes some of the intercurrent event as part of endpoint definition

• However clinical interpretation needs careful consideration
• Appropriate summary measure is required

– e.g., TTR Amyloidosis: Finkelstein-Schoenfeld (F-S) hierarchical ranking procedure and win ratio

52

Collaboration with Non-statistician

• Influence

– Become an indispensable member of the team by offering ideas of innovative designs, analyses, data summarization methods, and data presentation

• Relationship building

– Don’t bombard with statistical technical terminology – Educate team

• Effective communication

– Don’t raise problems and walk away. Offer a solution – Explain complex statistical concept in simple language

• Leadership

– Be a leader to drive end to end innovation – Understand the bigger picture

53

Conclusion

• Patients with rare diseases are in desperate need of innovation

– Requires a shift in thinking from 2 studies p<0.025 to continual learning via Bayesian approach

• Need to leverage ALL sources of information
• Meta-analytic approach provides great flexibility for borrowing
• Statisticians have a lot to add!

– “Fresh” perspective to study design – Perform “statistical engineering” for real life implementation – Train and influence non-statisticians

• An open-minded and collaborative attitude has been (and still is) the

most important factor

54

References

• Boxer AL, Lang AE, Grossman M, et al. (2014) Davunetide in patients

with progressive supranuclear palsy: a randomised, double-blind, placebo-controlled phase 2/3 trial. Lancet Neurol

• Gamalo-Siebers M, Hampson L, Kordy K, Weber S, Nelson RM,

Portman R (2019) Incorporating Innovative Techniques Toward Extrapolation and Efficient Pediatric Drug Development. Theraputic Innovation & Regulatory Science

• Golbe LI and Ohman-Strickland P (2007) A clinical rating scale for

progressive supranuclear palsy, Brain

55

References

• Neuenschwander B, Capkun-Niggli G, Branson M, Spiegelhalter D

(2010) Summarizing historical information on controls in clinical trials. Clinical Trials

• Neuenschwander B, Roychoudhury S, and Schmidli H (2016) On the

Use of Co-Data in Clinical Trials, Statistics in Biopharmaceutical Research

• Neuenschwander B, Roychoudhury S, and Branson M (2018) Predictive

Evidence Threshold Scaling: Does the Evidence Meet a Confirmatory Standard? Statistics in Biopharmaceutical Research

56

References

• Quintana M, Shrader J, Slota C, et al. (2018) Bayesian model of disease

progression in GNE myopathy. Statistics in Medicine

• Roy N. Tamura, Jeffrey P. Krischer, Christian Pagnoux, Robert

Micheletti, Peter C. Grayson, Yeh-Fong Chen, Peter A. Merkel (2016) A small n sequential multiple assignment randomized trial design for use in rare disease research, Contemporary Clinical Trials

• Satlin A, Wang J, Logovinsky V et al. (2016) Design of a Bayesian

adaptive phase 2 proof-of-concept trial for BAN2401, a putative disease- modifying monoclonal antibody for the treatment of Alzheimer's disease, Alzheimer's & Dementia. Translational Research & Clinical Interventions

57

References

• Schmidli H, Gsteiger S, Roychoudhury S, O’Hagan A, Spiegelhalter D,

Neuenschwander D (2014) Robust meta-analytic-predictive priors in clinical trials with historical control information. Biometrics

• Tolosa, E. , Litvan, I. , Höglinger, G. U., Burn, D. , Lees, A. , Andrés, M. V.,

Gómez‐Carrillo, B. , León, T. , and Ser, T. (2014), A phase 2 trial of the GSK‐3 inhibitor tideglusib in progressive supranuclear palsy. Mov Disord

• Wang G, Berry S, Xiong C, et al. (2018) A novel cognitive disease

progression model for clinical trials in autosomal-dominant Alzheimer’s

• disease. Statistics in Medicine
• Wei, B, Braun, TM, Tamura, RN, Kidwell, KM. A Bayesian analysis of small n

sequential multiple assignment randomized trials (snSMARTs). Statistics in Medicine.

58

Thank You