Ian David Lockhart Bogle and Michael Fairweather (Editors), Proceedings of the 22nd European Symposium on Computer Aided Process Engineering, 17-20 June 2012, London. 2012 Elsevier B.V. All rights reserved Biopharmaceutical Portfolio Management Optimization under Uncertainty Wenhao Nie, Yuhong Zhou, Ana Sofia Simaria, Suzanne S. Farid Department of Biochemical Engineering, University College London, Torrington Place, London, WC1E 7JE Abstract A stochastic combinatorial optimization decision-support tool is presented to address several interacting involved in portfolio management at both the portfolio level and the drug development process level. The tool comprises a genetic algorithm component, to search for the optimal solutions in the decision space, linked to an evaluation model of the drug development pathway that captures the interdependencies, value and risks of critical events for a portfolio of drugs. These components are both linked to an advanced database. The tool evaluates combinations of strategic by simulating the event flow of parallel projects using Monte Carlo simulations to generate probability distributions of the key profitability indicator, net present value (NPV). A tailored case study featuring the industrial development and commercialization of monoclonal antibody therapeutics was applied to validate the functionality of the tool. Potential strategies were evaluated based on the results of two objectives: the maximization of potential profit and the minimization of the probability of losing value. Scenarios are presented to highlight the impact of different portfolio related to drug candidate selection and out-licensing strategies on buildversus-buy capacity. The examples illustrate the benefits of using these techniques to explore large decision spaces effectively and investigate the interactions between portfolio and drug development on the risk and reward of strategies. Keywords: portfolio management, biopharmaceutical drug development pathway, combinatorial optimization, multi-criteria decision making, genetic algorithms 1. Introduction Biopharmaceutical drug development activities are highly costly, time-consuming and technology-intensive [1]. Decisions on capacity planning to supply material for clinical trials and the market are linked to decision on portfolio composition, portfolio scheduling and out-licensing strategies that also directly impact the financial capital available [2]. Furthermore the impact of pressures to adopt shortened clinical trial phase formats on manufacturing capacity need to be considered. The optimisation of strategic portfolio is complicated by the uncertain nature of drug development process including the duration and cost of clinical trials, the success/failure results of clinical stages, and the fluctuations in market sales. Software tools are essential to facilitating how best to invest resources for these multiple under uncertainty given the large decision spaces. Previous work has covered portfolio selection and project task scheduling using MILP [3] and build-versus-buy using brute force simulation [4] and genetic algorithm [2]. However, core related to out-licensing strategies and scheduling of facility builds have not been captured. This paper describes the development of a tool to optimize out-licensing strategies as well as candidate selection & build-versus-buy capacity under uncertainty. The tool
2 W.Nie et al. comprises a genetic algorithm linked to a detailed profitability model of pharmaceutical development lifecycle activities. 2. Methodology 2.1. Overview of the stochastic combinatorial optimization decision-support tool Fig.1 presents the main structure of the proposed framework of this stochastic combinatorial optimization decision-support tool. Starting from the left side of this figure, each individual solution is translated into including portfolio selection, project launch times, out-licensing strategy, clinical trial development format and manufacturing strategy. With these made, projects can be deterministically planned based on dependencies between necessary activities. Fig. 1 Structure of the stochastic combinatorial optimization decision-support tool. G(n) and G(n+1) refers to n th and n+1 th generation. NSGAII refers to non-dominated sorting genetic algorithm II. The tool performs Monte Carlo simulation on each solution to address the impact of uncertainties on its robustness as a portfolio management strategy. The tool generates stochastic inputs such as the cost and length of each activity, fluctuations in market sales and clinical trial results. Two financial indicators are derived from the stochastic simulation, namely the average NPV and the percentage that the NPV is positive for all instances. These two indicators are the main objectives to be optimized by this upport tool. A group of these solutions forms a generation in the context of the genetic algorithm. A non-dominated sorting genetic algorithm II (NSGAII) judges each individual solution on its merits in both objectives to decide its position in the next generation. Hence solutions evolve by generations and optimization is achieved. This tool was built in C# using Visual Studio 2008 (Microsoft Corporation, WA, USA) linked to a database in MySQL (Oracle Corporation, CA, USA). 2.2. Problem domain Biopharmaceutical portfolio management is characterized by several. On the portfolio level, the relate to portfolio candidate selection and out-licensing strategies. On the drug development process level, the decision variables focus on scheduling of each project s launch time and build-versus-buy capacity solutions across the development cycle stages. Key trade-offs to consider in portfolio selection are market potential, development costs and clinical trial risk. Out-licensing strategy opens the possibility that any product in the pipeline is acquired by a licensor in exchange for
Biopharmaceutical Portfolio Management Optimization under Uncertainty 3 immediate revenue and shared risk. Build-versus-buy capacity for manufacturing requires evaluation of trade-offs between higher investment costs for the build option versus higher operating costs when sourcing capacity from contract manufacturers. 2.3. Case study setup An industrially relevant case study is presented to illustrate how this tool can be helpful to a hypothetical biopharmaceutical company, with a given product candidate pool and limited manufacturing capacity, who wants to explore the potentials of various portfolio combinations and late-stage manufacturing options. This hypothetical company is focusing on novel therapeutic monoclonal antibodies (mabs). All product candidates are at the discovery stage. The company has a pilot scale facility for small-scale production and process development, but no large scale facility for late stage or commercial production. Table 1 presents the different categories of product candidates by their market potentials. Blockbuster products have the biggest market potential in terms of accumulated sales in 8 years, but at the same time are subject to high accumulated development costs and long development times, as well as high failure risk. The case study assumes 100% of blockbuster products falls into the category of high risk product. In contrast, niche products typically have much lower market sales but require lower costs and shorter development times, and are less risky. The case study assumes the same product has higher market potential in the hands of big pharmaceutical company (Pharma) than small biotech company (Bio) in that the former is known to have stronger marketing and distribution capabilities. The company has a product candidate pool of 1 blockbuster product, 4 medium products and 5 niche products. Table 1. Main characteristics of product candidates by category Acc. Sales Overall Product Acc. Cost ($million) development Category ($million) Pharma / Bio time (years) Blockbuster 20400 / 7837 192 16. 100 Medium 1781 / 898 144. 15 50 Niche 137 / 137 97 14 20 % of high risk product The case study setup of out-licensing deal terms comes from empirical data, adjusted by the following principles: 1) early deals result in the company retaining a larger value of the product than late deals; 2) deals with biotech companies result in the company retaining a larger value of the product than deals with big pharmaceutical companies. 3. Results and discussion The case study took approximately 450 minutes to run 100 generations of NSGAII, with each generation containing 50 individual solutions subjected to 1000 Monte Carlo simulations. Fig 2a presents the performance of solutions in generation 1, 17, 47 and 100 respectively in a case study scenario with on selecting product candidates and choosing the appropriate late-stage manufacturing strategy. Convergence towards the Pareto front occurs very obviously after 17 generations. Major improvement in the possibility of getting positive NPVs occurs between the 17 th and 47 th generation. From generation 47 to generation 100 there are minor improvements on both objectives in the
4 W.Nie et al. intermediate part of solutions to achieve a convex Pareto front. Fig 2b emphasizes progressions and newly emerged non-dominated solutions on Pareto front of the four chosen generations. Progression from generation 1 to generation 17 can be best observed within circle A where one solution from generation 1 is dominated by two solutions from generation 17. Similarly, progressions on other generations can be found in other circles (e.g. solutions from generation 17 in circle B to solutions from generation 47 in circle B). From Fig 2b there are quite a few overlapped dots which indicate that many solutions stay throughout generations. This is consistent with the setup of the optimization algorithm NSGAII where elitism is applied to preserve the best solutions in replacement of generations [5]. (a) (b) Fig. 2. Progression of GA solutions across different generations for (a) all solutions and (b) solutions on the Pareto front. Performances of solutions are measured by the possibility of achieving positive NPVs (x-axis) and the value of average NPV (y-axis). S_1, S_2 and S_3 are solutions chosen for detailed investigation of characteristics of the Pareto optimal front. A further investigation of the characteristics of the Pareto optimal solutions is revealed in Table 2. In this case study scenario, portfolio selection and late-stage manufacturing strategy are optimized to achieve maximum profit while balancing the potential risk. In Table 2, the low risk-low reward choice has 3 low risk products, whereas the others have only 2. The high risk-high reward choice has one blockbuster product, which on the upside could bring in large revenues, but on the downside costs more than average to develop. Therefore with regards to portfolio selection, the trend from high risk-high reward to low risk-low reward strategies can be linked to the selection of more products with high development costs but large market potentials to low development costs with low market potentials. As for the late-stage manufacturing strategy, the high risk-high reward solutions choose to build facilities for all products but one, although the construction time is as late as the case study setup permits. The low risk-low reward solutions only build facilities for 2-3 of their products and outsource the remaining manufacturing requirements to a contract manufacturer. Similar characteristics were found for neighbouring solutions in each region of the Pareto front.
Biopharmaceutical Portfolio Management Optimization under Uncertainty 5 Table 2. Portfolio selection and late-stage manufacturing decision of Pareto optimal solutions of various performances Solution Decision Drug 1 Drug 2 Drug 3 Drug 4 Drug 5 variable High risk-high Portfolio Niche Medium Niche Blockbuster Medium S_1 on Fig 2b Build Late++ No build Late+ Late++ Late++ Medium riskmedium Portfolio Niche Niche Medium Niche Medium S_2 on Fig 2b Build No build No build Late+ Late Late++ Low risk-low Portfolio Medium Niche Niche Medium Niche S_3 on Fig 2b Build No build Early+ Early No build No build Note: Decision variables displayed in this table are from Pareto optimal solutions of 100 th generation. The use of Niche, Medium and Blockbuster refers to the category that product candidate falls into. Background darkness indicates the risk profile of product: white background low risk; grey background medium risk; black background high risk. The use of Late and Early in build refers to the timing of construction of large-scale manufacturing plant. 4. Conclusion The stochastic combinatorial optimization decision-support tool produces a series of optimized solutions for the decision maker to choose from according to his own riskreward preference. The tool provides insights into the factors influencing financing of biopharmaceutical enterprises. It is flexible to accommodate new and uncertainties, robust to withstand extensive data exchange between the model and the database during the GA procedure, and efficient in running time. 5. Acknowledgements Data support from Deloitte Recap is gratefully acknowledged. References [1] J. DiMasi and R.W. Hansen, 2003, The price of innovation: new estimates of drug development costs, J Health Econ. 2003 Mar;22(2), 151-85 [2] E.D.George and S.S.Farid, 2008, Stochastic Combinatorial Optimization Approach to Biopharmaceutical Portfolio Management, Ind. Eng. Chem. Res., 2008 47(22), 8762-8774 [3] D. Subramanian, J. F. Pekny, G. V. Reklaitis and G. E. Glau, 2003, Simulation-optimization framework for stochastic optimization of R&D pipeline management, AIChE Journal, Volume 49, Issue 1, pages 96 112 [4] A. Rajapakse, N. J. Titchener-Hooker, S.S. Farid, 2006. Integrated approach to improving the value potential of biopharmaceutical R&D portfolios while mitigating risk. Journal of Chemical Technology and Biotechnology 81, 1705 1714. [5]K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, 2002, A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation, April 2002, Vol.6, No.2, 182-197