The Role of Literature Models to Support Strategic Decision-Making in Clinical Drug Development
by William Ebling, PhD, Jill Fredrickson, PhD, and Adam Rutkin
The development of a new drug is a high risk and costly activity that involves multiple decisions and tradeoffs. This occurs in an environment of limited and imperfect data. Modeling and simulation techniques provide a mechanism for capturing the current knowledge and associated uncertainty about a new compound relative to its competition, and provide a framework for enhancing strategic decision-making.
Model-based approaches for combining proprietary data on a specific compound with literature data of its likely competitors offer powerful tools for managing risks concerning dose selection, efficient and effective trial design and key "go/no-go" decisions. As part of its Critical Path Initiative, the FDA is developing an internal library of therapeutically-focused drug-disease models based on literature data and its own vast collection of clinical trial data from sponsor submissions, and is considering how to begin sharing these models with industry.
The benefits of integrated, comparative literature modeling include:
- The ability to make explicit quantitative assessments of competing treatments even when no direct within-study comparisons are available
- The ability to predict clinical outcomes for NCEs even where no clinical data are yet observed (e.g., at unstudied doses or the extrapolation of steady-state effect from early short-duration dose ranging trials)
- The ability to bridge preclinical-to-clinical or biomarker-to-clinical relationships within a common therapeutic class
- Greater objectivity when predicting and inferring an NCE’s critical efficacy/safety/tolerability attributes (e.g., comparison of NCE attributes and uncertainty with competing therapies and draft launch label targets)
These benefits mean higher drug development productivity, improved clinical quality and commercial performance of new therapies, as well as greater transparency for decision-making throughout the development team and across the organization.
Literature Modeling Approach
Building models of the competitive landscape for a new drug using public-domain scientific literature can be started well before patient-level data for the new compound becomes available. The studies in the available literature likely have different designs, and thus may have generated data that are difficult to compare. Integrated literature/NCE models provide a way to pool such heterogeneous data, supported by statistical techniques that take special account of variability among outcomes measurements, patient characteristics and trial designs to quantify variations and uncertainty of drug effects. In some cases, models based on summary data (e.g., population mean) as a function of dose may be sufficient to provide insights for dose selection or competitive positioning. For other model-based applications, such as simulating clinical trials, a model capable of simulating individual patient responses may be required.
Prior to building the model, the clinical development team defines important drug attributes that impact critical development questions. The type of model and data required depends on how the model will be used. Once a modeling strategy is developed, the data necessary for constructing the proposed model and viewed as convincing by the decision makers is extracted from the literature and placed in a database. Relevant public-domain data sources include:
- Published journal articles
- Regulatory documents (e.g., FDA Summary Basis for Approval)
- Package inserts or promotional materials from the drug manufacturer
- Published abstracts or poster presentations
- Web documents (e.g., press releases about new clinical trial results)
- Online clinical trial registries
Proprietary in-house data on the new drug can be added to the public-source database and used to co-model the relevant attributes of the new drug and its competitors. Combining summary-level literature data and proprietary patient-level data require careful attention. Issues related to model selection (e.g., high order mixed-effects models may be required) and to incorporating relevant covariates or uncertainty components of the model must be considered. For example, random effects may include inter-trial and inter-arm variability as well as inter-patient and residual variability. In addition, it is important to recognize that model predictions based on data available in the public literature may be over-optimistic, given the inherent publication bias against negative or equivocal results. Moreover, published studies may represent historical results in a disease population that do not fully represent the current target population due to the evolution of drug therapies and clinical practice over time (for further discussion of these issues, refer to Model-Based Meta-Analysis for Integration of Data from Multiple Sources).
After the models have been developed, they are validated and qualified in order to ensure that the development team will have a high degree of confidence in the predictions and insights yielded. The models can be then used to simulate the current understanding of critical dose-endpoint relationships and the impact of relevant covariates on these relationships. Expected dose response relationships for the NCE can be contrasted with the competition and draft launch label targets to objectively select doses and understand the value of the NCE at its optimal dose and target population. In addition to understanding the attributes of the NCE, the model can be used as a basis for clinical trial simulation to understand design characteristics that will be required to demonstrate the compound's clinical value.
The end result is a model that codifies the most important aspects of the competitive landscape, given a current understanding of the new drug and its position within that landscape. The model serves as a transparent representation of information that is useful for making high value strategic decisions, and can be updated as new information becomes available, serving as an evolving base of knowledge throughout the development cycle.
Case Examples
Competitive Profiling to Support a ‘Go/No-Go’ Decision. A recently published example of the use of literature modeling to support quantitative decision-making in early development is provided by gemcabene, a novel cholesterol-modifying agent developed by Pfizer [Mandema et al. AAPS Journal. 2005; 7(3)]. Pfizer undertook an integrated PK/PD modeling and simulation strategy to evaluate the likely clinical profile for gemcabene based on all relevant data, including prior knowledge on competing treatments, which included marketed statins, the cholesterol absorption inhibitor ezetimibe, and their combinations. The Pfizer team was particularly interested in whether gemcabene in combination with statins could compete with the established ezetimibe-statin combination therapy for patients with high cholesterol.
In the first phase of the modeling effort the gemcabene team, in collaboration with scientists at Pharsight Corporation, performed a literature meta-analysis while a small Phase II gemcabene-plus-statin trial was in progress. The analysis combined publicly available, summary-level competitor data from 21 randomized trials, spanning more than 9,000 patients, with internal patient-level data for gemcabene from earlier Pfizer studies. A variety of response endpoints were modeled, with the primary purpose of providing a quantitative framework to assess gemcabene’s ability to lower “bad” (LDL-C) cholesterol versus the competition. Based on these data, models of several response endpoints were built and tested for gemcabene and six competitors. The team then used the model to simulate virtual product profiles across multiple competing therapies, efficacy and safety endpoints, and patient characteristics. The goal was to use this database to explore the complex domain of product trade-offs in advance of receiving the Phase II gemcabene trial results, and to quickly update the models when this new data became available to more fully inform decision-making. Pharsight's Drug Model Explorer™ visualization software was used throughout the project to make modeling results readily available to the team and to communicate predictive results of modeled product profiles.
When the Phase II trial results were incorporated into the analysis, the resultant comparative models predicted that the gemcabene-statin combination would not provide a superior LDL-C lowering benefit relative to the ezetimibe-statin combination therapies over the relevant dose range. Development was discontinued, avoiding further direct investment of approximately $2 million and four to six months to perform an additional trial1. The literature-backed modeling strategy contributed significantly to this decision, providing a quantitative assessment between treatment options that were not directly compared in the Phase II trial. The integrated modeling strategy also increased the certainty of the decision to stop development.
Use of Biomarker and Literature Data. There are many situations where a clinical endpoint of interest has a relatively long time course, but related biomarkers with shorter time courses of action are available and may be utilized to speed drug development. For example, in Type 2 diabetes mellitus glycosolated hemoglobin (HbA1c) is typically the primary endpoint of interest, while fasting blood glucose (FBG) is often used as an early biomarker for treatment efficacy. Though both are affected by drug therapy, HbA1c measures inherent change more slowly than glucose measures due to the relatively long life span of hemoglobin-carrying red blood cells.
In this case, literature-based disease models that describe the relationship of FBG and HbA1c as a function of time can be built using data from currently marketed compounds as monotherapy, combination therapy, or replacement therapy. Other valuable insights into the treatment of Type 2 diabetes can also be gleaned from the literature models, such as the mean time course of disease progression, or the variability in the mean placebo response seen in trials. The literature-based disease models can then be combined with FBG response data from an NCE's short-term trials. The combined model can be used to make predictions about the steady-state HbA1c response to facilitate the design of pivotal trials for the NCE, along with providing comparisons with currently approved treatments.
Trial Design Optimization. Models built using literature data can be readily applied to optimize trial design via clinical trial simulation. In one blinded development example, the Phase II objective for an NCE was to efficiently find a non-inferior dose relative to the standard of care with sufficient certainty to carry a single dose into Phase III pivotal trials. The modeling strategy combined pre-clinical and Phase I data for the NCE with public-source clinical data on five competing drugs, representing 27 clinical trials and more than 75 treatment arms. This approach enabled construction of a dose-response model to predict clinical outcomes for the NCE. The model was then used to simulate alternate clinical trial designs.
The trial simulations enabled efficient design and analysis of the Phase II trial to select a dose for Phase III. The range, number and spacing of doses was optimized, and the design included adaptive “pruning” strategies to assign subjects to the most relevant doses. The Bayesian modeling and analysis methods underlying this approach enabled a shorter and more informative Phase II program compared to conventional designs that did not fully incorporate prior information from the trials literature – a reduction amounting to hundreds of subjects2. The literature-based dose finding strategy also shortened time-to-market by at least four months, and potentially much more, by avoiding incorrect dose selection and the higher likelihood of a failed Phase III program.
Outlook
FDA has a growing base of experience with literature modeling and model-based drug development, as part of its Critical Path Initiative. Dr. Robert Powell, director of FDA’s Division of Pharmacometrics, recently mentioned that for an HIV model FDA combined the published work of a prominent New York AIDS clinic with work from experts on viral dynamics at Los Alamos National Laboratory3. The same model has since been used by a number of pharmaceutical companies in their HIV development programs.
Beyond HIV, FDA includes relevant literature estimates for comparative drug effects as a deliverable in its voluntary End of Phase IIa meeting. The purpose of this meeting with selected sponsors is to reduce unnecessary failures in late stage clinical trials through nonbinding, scientific review of modeled exposure-response data from preclinical and clinical trials. FDA also continues to develop an internal library of therapeutically-focused drug-disease models (e.g., in diabetes, Parkinson's) based on the literature and its own vast collection of clinical trial data from prior NDAs and SNDAs. Proprietary source data from sponsors used to develop and validate the models would not be shared publicly, but FDA is considering how to begin sharing the models with industry4.
Predictive models of the competitive landscape for a potential drug, based on public domain scientific literature, help provide a quantitative assessment of likely clinical benefits and trade-offs to support decision-making. Ideally, these models should be introduced early in development and can be rapidly updated as new information on the NCE becomes available. Early experience with these models provides the development team time for input in the model building and qualification process, and an opportunity for all members of the team to gain confidence with interpreting and communicating model-based representations of drug effect and uncertainty. Public-source literature models, particularly when combined with proprietary data on an NCE of interest to produce an evolving picture of a compound's competitive product profile, offer a powerful and flexible set of tools to support quantitative decision-making in clinical drug development.
1PK/PD Simulation Speeds Decision Making. Bio-IT World Best Practices [serial online]. January 23, 2006.
2Gillespie B. Model-Based Meta-Analysis for Integration of Data from Multiple Sources. Presented at: Pharsight Corporation Modeling and Simulation Web Conference, February 15, 2006.
3Russell J. FDA Mulls Drug/Disease Model Library. Bio-IT World [serial online]. December 19, 2005.
4Powell R. FDA Experience with End of Phase IIa Meeting: An Attempt to Improve Drug Development Decisions. Presented at Clinical Pharmacology Subcommittee of the Advisory Committee for Pharmaceutical Sciences. November 14, 2005.