Table 5.6 specifies the key uncertainties considered, together with estimates of their discrete probability distributions. These values were determined from literature sources and through discussions with industrial experts (eg. D. Sherwood, Lonza Biologies, Slough, England).
Table 5.6 Key risk factors and their probability distributions
Risk factor Possible values Probability
Product titre 0.10 g/L 0.1 0.25 g/L 0.2 0.40 g/L 0.4 0.55 g/L 0.2 0.70 g/L 0.1 Product demand 15g 0.1 30g 0.2 45g 0.4 60g 0.2 75g 0.1
Market success true 0.25
false 0.75
The titre defines the grams of product expressed per litre of fermentation broth and significantly influences the manufacturing costs as it determines the number of batches required to satisfy the product demand. Typical antibody titres in batch
APPLICATION OF Si mBi o p h a r m a
cultures are 0.1-0.5g/L (Walsh, 1998). However, higher antibody titres in the order of 1-3 grams per litre have been reported using fed-batch cultures of mammalian cells (Birch, et a l, 1995; Bibila & Robinson, 1995; Xie & Wang, 1996). Six representative titres were therefore selected to represent typical titres at Phase I clinical trials, considering that titres in the order of grams per litre were unlikely at this stage since they require optimisation of feeding strategies which may not be a priority at such an early stage of development. The titres were assumed to follow a discrete normal distribution as indicated in Table 5.6.
Demands for Phase I clinical trials vary from milligram quantities to tens of grams (Walsh, 1998) depending on the cumulative dose of the drug and the number of patients in the trial. The amount of product that needs to be manufactured must take account of both kinds of use that will be made of the drug candidate. Non-clinical usage addresses quality control and validation issues, such as the need for in-process samples, release and stability samples, and retain samples. Projections of investigational product requirements can be made assuming a 25-300% excess over the actual subject usage is necessary for non-clinical uses (Bernstein & Hamrell, 2000). The values selected for the manufacturing product demands were assumed to follow a discrete normal distribution.
The likelihood of a drug candidate never breaching the market is also considered owing to the high risk of failure in clinical trials. Sofer & Hagel (1997) and Breggar (1996) commented that only 23% of drugs entering clinical trials became marketed drugs. Other authors have provided slightly more optimistic success rates from Phase I to launch of 34% (Mackler & Gamerman, 1996) and 67% (Struck, 1994). For this analysis the probability of market success was assumed to be 25%. However, more recent data suggest that the success rates for biopharmaceuticals are possibly much lower, eg. Avgerinos (1999) suggests a typical value of 15%. As mentioned earlier, the primary aim of this study was to demonstrate the functionality of the approach designed in this thesis and implemented in SimBi o p h a r m a. Hence, the actual value
used for market success probability is not of critical concern. Assuming an appropriate order of magnitude estimate provides the basis for a hypothetical case focusing on how to map manufacturing data into a simulation model and interpret the simulation results.
APPLICATION OF Si mBi o p h a r m a
Having identified the major technical and commercial uncertainties to be incorporated into the analysis, the models set up in the prototype tool were then modified to capture these and perform stochastic modelling using Monte Carlo simulations. Additional inputs included the probability distributions of the uncertain factors (Table 5.6), the length of each simulation and the number of simulations required to reach convergence in the output distributions. The probability distributions were captured using branch blocks, as described in Chapter 4 (eg. Figure 4.7). This resembled a decision tree approach to representing the possible values for each uncertain parameter. The risk parameter initialisation block indicated in Figure 5.2 contained these branch blocks on its subworkspace and ensured that a fermentation titre and product demand were randomly selected prior to commencing a manufacturing campaign. At the end of each campaign the success of the candidate in reaching the market was predicted in the detail of the block labelled “update-economics” (Figure 5.2). The Monte Carlo simulation technique was used to run several simulations and hence determine the resulting frequency distributions of the output summary measures. This enabled the investment options to be assessed based on both the expected outputs and their associated risks.
Each run simulated 48 weeks and the key performance measures were the number of project campaigns pushed through the plant, the number of drug candidates that reached the market and the cost of goods per gram. Having validated the results of a single simulation in the deterministic analysis, 300 simulation runs were performed to characterise the variability in the performance measures due to uncertainties in the product demand, titre and market success for each case. To determine the number of simulation runs required to reach convergence, running averages of the results were monitored until they levelled off. Frequency distributions of the performance measures were generated and various statistics computed to aid the decision-making process. The following discussion highlights possible analyses that can be performed in order to manipulate and interpret the data from the Monte Carlo simulations and to evaluate manufacturing options.
APPLICATION O F Si mBi o p h a r m a
5.6.2 Base Monte Carlo simulation results