Parnell et al [29] define a system boundary as a physical or conceptual boundary that contains all the essential elements, subsystems, and interactions necessary to address a systems decision problem. Different decision objectives would motivate different definitions of systems boundaries.

We generalized the systems in the eligible articles into two types: (1) product development systems, and (2) manufacturing and supply chain systems.

All the articles we identified focused on cell and ex vivo gene therapy; we found no decisional tools in tissue engineering or in vivo gene therapy. Table 2 [19-27] shows the product types and type of transplant considered in the articles.

There was considerably more focus on allogeneic therapies and mesenchymal stem cells (MSCs) as the cell source. The study by Weil et al [24] is the only one in this review that considered autologous processes, but it focused only on downstream affinity purification.

Decision variables are variables that the decision-maker controls. Such variables are dependent on the decision objectives and the problems they seek to answer.

For product development systems, the models seek to answer to objectives such as minimizing development duration, risks, and investment costs. The timing for technology change was the decision variable modeled by Hassan et al to study the impact of process change in product development on the project NPV [21]. McCall [27] studied the impacts of product development risks and uncertainties, as well as rework probability on the investment costs; as there was no optimization module in this study, no decision variable was identified.

For manufacturing and supply chain systems, the models seek to answer to objectives such as process yield [19] and manufacturing cost minimization [22,24-26]. In these systems, the choice of unit operations in the manufacturing process sequence, chosen equipment, and their capacities are critical decision variables to be considered. Figure 4 shows the upstream and downstream unit operations options considered in the articles. Together with process parameters, process flowsheet, and requirements, the processes were modeled in order to understand the associated cost of goods. Conclusions on the relative advantages of various technology options under different demand scenarios were drawn for bottleneck analysis and decision recommendations.

Input parameters are defined as “a constant element or factor, especially serving as a limit or boundary” [31]. These are inputs defined as prerequisites of the objective function; in other words, they are constants of the objective function and not the values to be optimized. In the 9 articles we identified, these included scale, throughput, demand goal, and technical process parameters.

Early articles in the area generally looked at demands several times more than the recent articles. As more commercial case studies of regenerative medicine products arise, the demand landscape is better understood and the estimation for cost-of-goods modeling has been lowered from the monoclonal antibodies ballpark (1000-500,000 doses of allogeneic MSCs per year [22] to around 2500 doses/y for a regional center for allogeneic MSCs [25] and 500-5000 doses/y for chimeric antigen receptor T cells (CAR-T) [26]). Demand scenarios were chosen depending on the cell type and the therapeutic target. As more and more real-world commercial cases have emerged, recent articles gave a lot more consideration to the clinical applications and their specific demands and, hence, proposed more realistic demand scenarios.

For upstream operations, comparisons were drawn from a range of equipment scale, cell culture modes, and extent of automation. Table 3 [19,22,25,26] summarizes process parameters previously considered and explicitly mentioned in their respective articles.

In all the articles, planar culture flasks (eg, T-flasks and multilayer flasks) were consistently found to be the most expensive of all evaluated technologies for allogeneic therapies and infeasible for higher cell number per lot. The number of cells harvested per surface area was found to be the most important cost driver, as it dictates the number of expansion units required and, hence, the raw materials and consumables requirements [22,26].

Table 4 [23,24,26] shows the process parameters for downstream processing discussed in the articles.

The downstream process starts with the wash and concentration step, and common wash concentration unit operations were discussed in detail by Hassan et al [23] and Jenkins and Farid [26]. Hassan et al reported that wash and concentration downstream steps were a bottleneck for high-cell-dose lots at high demand. As the demand estimate was lowered in the study of Jenkins and Farid, wash and concentration was no longer shown to be a bottleneck except in integrated bioprocess platforms such as CliniMACS Prodigy, which has a relatively smaller volume-reduction capacity.

Following wash and concentration, affinity purification has also been a target for modeling and optimization. Weil et al [24] and Jenkins and Farid [26] looked into affinity purification for autologous induced pluripotent stem cells and allogeneic CAR-T cells, respectively. Weil et al compared fluorescence-activated cell sorting (FACS) versus magnetic-activated cell sorting (MACS) and evaluated a novel technology that does not require cell labeling. MACS was determined to be more cost effective for dose sizes with a higher cell count (>7.0×10⁷ cells/dose), as FACS is limited by its process scale. The model by Jenkins and Farid [26], for affinity purification, considered only MACS and integrated bioprocess platform.

The articles made other assumptions besides technology-related assumptions and constraints.

For scheduling-related problems where one task follows another, task precedence constraint was used. McCall [27] used the task precedence constraint for dictating the start and end of a task, which is useful for setting up the scheduling problem. Iteration loops were built in the development pathway with assumption of learning. Hassan et al [21] assumed a linear project development pathway with failure probability. They constructed a database with information on clinical trial development times and failure rates of all 592 commercial cell therapy projects from 1981 to 2011 to estimate the duration and failure rate of similar products. This approach allows more industrially relevant benchmark assumptions to be made and, hence, gives rise to higher-quality results.

For resource utilization, McCall [27] assumed a fixed and steady-rate consumption of resources and a renewable resource pool throughout the project duration. Similar assumptions were made in all the other cost models to better understand the impact of resource utilization on the overall cost of goods. For instance, Simaria et al [22] showed that efficient use of equipment and facility can lower the depreciation costs shared across doses. Harrison et al [25] looked into the impact of human resource turnover in detail to understand the impact of increased operators on the relative cost of labor in overall cost of goods.

Having reasonable cost assumptions is one of the most important factors determining the validity of the model. Table 5 [22-26] shows some of the cost assumptions used in various models.

The cost assumptions of Harrison et al [25] were considerably more detailed than those in the rest of the articles. Quality control test panels were in accordance with good manufacturing practice requirements of the specific product and are listed in detail in the supplementary material of Harrison et al and based on industry information from Athersys. Depending on product characteristics, each product needs different quality control tests and assays requirements and, hence, the costs can be quite different. For instance, genetically modified cells would require assays on transformed cell populations to demonstrate appropriate and reproducible expression of newly acquired characteristics [25].

Labor costs can be quite different depending on geographical location. In an interview with the chief executive of Nanjing Legend, a Chinese company, he estimated that the manufacturing costs for CAR-T in China can be one-sixth of those in the United States due to cheaper overheads [32]. Simaria et al [22] suggested in their sensitivity analysis that labor rate is one of the most important cost drivers for less-automated processes.

The two main methods for accounting for facility costs in the studies were equipment-factored estimates (eg, Lang factor) and estimates of cost per square meter. Facility costs are averaged over the period of depreciation and shared among all doses. The Lang factor is a commonly used method in project cost estimates in the engineering industry and is recommended by the American Association of Cost Engineers [33]. The Lang factor used by Hassan et al [23] was taken from Pollock et al [12], which took into account pipework and installation, process control, instrumentation, electrical power, building works, detail engineering, construction and site management, commissioning, and contingency factor. It was unclear what Harrison et al [25] included in their method of cost per square meter applied to cleanroom space. In addition, a different cleanroom grade would constitute a different cleanroom space running cost and, hence, it is important to understand the good manufacturing practice requirements of the manufacturing environment.

To ensure all relevant costs are identified, typically models simulating the actual manufacturing or product development process are constructed and costs associated with each step are summed. Costs were analyzed by Simaria et al [22], Hassan et al [23], Hassan et al [21], Weil et al [24], Harrison et al [25], and Jenkins and Farid [26]. This method allows for process-centric costing, which in turn supports cost analyses based on technology options.

A value system modeling is a way of modeling the firms by sets of activities that the firms use to create value and competitive advantage [34]. McCall [27] modeled the set of activities in product development and accounted for development process characteristics such as interdependency, iteration, activity cost, and duration uncertainties. Through this model, McCall was able to highlight the critical processes, resources, and risks in product development. The report highlighted the importance of early-stage investment, clinical trials rework, and regulatory requirements.

Design structure matrix is a method developed by Steward and other in 1981 for planning and communicating engineering works [35]. The matrix represents the events, their sequence, and the interdependencies between events. McCall [27] used a design structure matrix to clearly represent the precedence constraints while considering iteration circuits inherent in product development.

Several articles included what-if scenario analysis, where the dose sizes, lot sizes, and demand for products were varied. These studies were used to provide guidance for technology selection under different circumstances.

While most articles dealt only with manufacturing or investment cost optimization, Jenkins and Farid’s model employed a multicriteria decision-making methodology to assess bioprocess flowsheets [26].

The weighted sum technique provided a way to account for both quantitative and qualitative attributes of a solution, and, by assigning weightings, considered the perceived relative importance of different attributes. Weighted sum, however, is just one of many methods of multi-attribute decision making. Hester and Velasquez [36] conducted a comprehensive review and comparison of the methods commonly used. The analytic hierarchy process allows for pairwise comparisons to compare alternatives, which is less data intensive and more suitable for qualitative performance-type problems and resource management applications.

Common themes incorporated into these manufacturing and development cost models are the risks and uncertainties lurking in the industry. The major methods of capturing risks and uncertainties in the studied models were stochastic modeling, Latin hypercube sampling, Monte Carlo analyses, and sensitivity analyses.

Deterministic models use discrete values, meaning that, for a certain input, the output will always be the same. Stochastic models have at least one quantity with random values, leading to an ensemble of different outputs [27].

Stochasticity was accounted for in 3 of the articles, where triangular distributions were applied to parameters to capture the uncertain and variable nature of the systems. McCall [27] applied a triangular probability distribution to the task duration to capture the uncertainties in development step duration. Hassan et al [21] applied a probability distribution to the success rate of each development step. Jenkins and Farid [26] assigned probability distributions to the weighting of quality attributes in the multi-attribute decision-making module; hence, they modeled only the variability in preference of quality attribute.

We noted that probability distribution had not been assigned to variability and uncertainties in the manufacturing process in these studies.

McCall [27] categorized the risks into product risk factors and enterprise risk factors. Product risks were defined as risks that can harm the patient, namely the choice of cell type, manufacturing processes, and delivery mechanism. Enterprise risks were defined as risks that affect the commercialization of the product and the business developing the product, namely technical risks and market risks. The Latin hypercube sampling method was used to consider the probability of failure and duration for each task along the product development pathway and the interdependencies. It is worthwhile to note that in this model, iteration caused by failures and impact of failures during each phase were considered using three matrices: design structure, rework probability, and rework impact.

Hassan et al [21] simulated the risks and uncertainties of process change along the product development pathway through Monte Carlo analyses. To adjust the project NPV according to risk, they used a discount rate based on the riskiness and expected development time.

Sensitivity analysis is a common strategy to account for uncertainties and identify key cost drivers [22,24,26]. It is useful in understanding which assumptions and parameters the overall system is most sensitive to. Simaria et al [22] focused on upstream production for MSCs and found that the main cost drivers were microcarrier area, harvest density, media price, and downstream yield. Weil et al [24] and Jenkins and Farid [26], who considered processes requiring differentiation or gene modification, consistently cited the key cost drivers to be the efficiency of differentiation and gene modification. This is particularly interesting, as Weil et al modeled an autologous process [24] while Jenkins and Farid modeled an allogeneic process with a different cell type [26].

Hypothetical case studies where the demand and dose sizes varied were conducted for manufacturing systems, including by Simaria et al [22], Hassan et al [23], Weil et al [24], and Jenkins and Farid [26]. The case studies were useful in the evaluation of process bottlenecks and technology-switch sweet spot analysis.

Figure 5 shows the simulation platforms employed in the different tools. Depending on the goal of the decisional tool, different simulation tools have been used. For simpler models, using Microsoft Excel with Visual Basic has appeared to be sufficient. For instance, the Excel model constructed by Weil et al [24] consisted of a cost model with mass balance, design, sizing, resource utilization and cost-of-goods equations, database of bioprocess technology data and iterative algorithms, and scenario analysis developed using Visual Basic. However, Visual Basic codes are susceptible to Excel program upgrades, and changing formats (eg, adding a column or a row) may cause changes in the functions. C# or MATLAB allows more versatile coding experience, and for models requiring many runs, such as uncertainty or stochasticity analysis, these platforms may be more suitable.

For models with larger databases, it is worth looking into database software such as Microsoft Access. Database software provides better scalability if the volume of data is huge, and the links can be built in a more robust ways than in spreadsheets.

Visualization software such as Google Charts allows for information to be easily updated and visualized and hence is useful for presenting a lot of data in a meaningful way [20]. Dedicated add-ons, such as Palisade Risk 6, used by McCall [27], allow for Monte Carlo simulations and sensitivity analysis to be carried out a bit more easily. Without experience of using the software, however, the usability of this software has not been formally compared with implementations based on C# or MATLAB.