Introduction
New antibiotics are not being developed fast enough to offset the threat from antimicrobial resistance. Antibiotics are not sufficiently profitable for developers and investors to engage. This translates into organizations either (1) not engaging in discovery so that no new projects enter the clinical pipeline, and/or (2) prematurely discontinuing projects that are already in the pipeline. In attempts to mitigate the combination of increasing resistance and decreasing antibiotics R&D, leading scientists and policy advisors have called for policy interventions that incentivize pharmaceutical organizations.Reference Cooper and Shlaes1
Gauging how different designs of an intervention might result in different numbers and types of antibiotics is paramount when balancing billions of dollars of public money. This requires understanding the impact of interventions on pharmaceutical organizations' propensity to engage in antibiotics R&D. Thus, to model the efficacy of any policy intervention, we must explore: (a) how often a new antibiotic enters the pipeline, (b) under what conditions, and by whom this antibiotic is further developed or discontinued, and (c) how a policy intervention may be introduced into this system. This work is concerned with the two latter points.
Previous works in modeling the effects of policy on antibiotics R&D have taken a macro- or health-economic perspective, without accounting for agent heterogeneity.Reference Sertkaya, Eyraud, Birkenbach, Franz, Ackerley, Overton and Outterson2 However, it has been shown that explicitly considering heterogeneity may yield qualitatively different policy recommendations than when exploring the dynamics of “average” agent characteristics.Reference Dawid and Neugart3 Therefore, this work explores the effects of the policy intervention known as Market Entry Reward (MER) for antibiotics R&D using agent-based simulation, which explicitly considers the heterogeneity of economic agents.
Large pharmaceutical organizations have left the antibiotics space as it is insufficiently profitable4 when compared to other therapeutic areas. Long development times and low probability of success means that rational investors must expect sales well beyond R&D costs, which in the case of antibiotics is not realistic.5
Monetary policy interventions can broadly be categorized as being either push or pull based. A push intervention supports an activity by providing upfront funds to undertake it, while pull rewards actors only after an achieved goal. In pull schemes the risk is thus carried by the receiver and in push schemes by the payer. Push lowers the cost of entry, while pull increases returns.Reference Grace and Kyle6 A MER is a classical pull intervention comprised of a monetary reward paid to an actor upon successfully bringing an eligible drug to market.
We consider two variations of MERs: fully delinked (FD) and partially delinked (PD), both paid over five years directly after approval. Delinkage is the notion of detaching a product's revenues from its sales volume, and may be achieved by offering a fully delinked MER.Reference Rex and Outterson7 A partial substitution of revenues from sales volumes is achieved by a partially delinked MER. An FD MER entails a payment provided as an alternative to market sales (i.e., replaces them), while a PD MER entails a payment provided in addition to market sales (i.e., supplements them).
Figure 1 illustrates the cash flow over the lifecycle of an antibiotic where R&D costs are followed by either a partially (left) or a fully delinked MER (right). In both cases, the MER is paid over five years. Both FD and PD MERs have the potential of turning loss-making projects into profit-making ones.
Depiction of a partially (left) and fully (right) delinked market entry reward in the context of our simulated development costs and global net revenues. Lines show the full range of input data for every cash flow.

The research questions that we explore in this paper are (1) how the size of a fully and partially delinked MER influences the likelihood of market approval of new antibiotics and (2) how these approvals vary depending on changes in key economic parameters such as forecasted product sales, R&D costs, and investors' expected profitability levels.
We have constructed an agent-based Monte Carlo model, with input vetted by stakeholders of the multidisciplinary and multi-stakeholder project DRIVE-AB, to simulate antibiotic projects from discovery through to sales. This R&D process is complicated, and involves successfully completing trials, acquiring funding, and potential changes in ownership structure. Treating and studying antibiotic projects individually allows us to assess the impact of interventions to stimulate the R&D pipeline of antibiotics as a whole. By means of sensitivity analysis, we explore the effect of four key parameters involved in R&D investment decisions: projected net revenues, R&D costs, venture capitalists discount rates, and large pharmaceutical organizations' financial thresholds.
We confirm that pull-based interventions effectively stimulate antibiotics R&D since revenues have the strongest effect on the likelihood of an antibiotic reaching market.
Method
Assessing the effects of MERs on antibiotics R&D requires consideration of a large number of possible scenarios for a given set of inputs, each one unfolding during one “run” of the simulator. With tens of thousands of scenarios per experiment, it is not feasible to construct and amalgamate the outcomes by hand, which naturally leads to the adoption of computer-based simulations.
We have constructed an agent-based Monte Carlo simulation that explores the likelihood of antibiotics reaching market under different initial conditions (technical probability of success; development costs; development times; expected revenues; requirements of venture capitalists, partners, and acquirers) and under different MERs.
In our simulation, an antibiotic's likelihood of reaching market approval (LOMA) is determined by (1) the risk of technical failure at the various stages and (2) the financial decisions made by developers. We calculate LOMA as m/n where m is the number of antibiotics that reached market, and n the number of antibiotics that entered pre-clinical development.
Antibiotic projects enter the simulation in pre-clinical at an artificially derived rate, which we term the “entry rate.” The technical success of an antibiotic in each phase is assumed to follow reported standard rates and is assigned in a purely probabilistic manner using triangular distributions.
We assume that, for projects that do not encounter technical failure, developers and financiers decide whether or not to provide and maintain funding based on Expected Net Present Value (ENPV), a widely used metric to model decision-making in pharmaceutical organizations.Reference Blau, Pekny, Varma and Bunch8 We calculate ENPV as:

where i is the discount rate (cost of capital) of the evaluator, n the final month of cashflow (in our simulation, usually patent expiry), P(n) the probability of reaching the final month, and C(t) the cashflow at time step (month) t.
We assume that projects are evaluated on the basis of their ENPV only before transitioning to a new stage, hereon referred to as a “decision-point.” A project is pursued if ENPV is greater than or equal to the threshold of the actor investing in the project, and is otherwise TERMINATED.
While all projects enter the simulation in pre-clinical, they leave the simulation either as TERMINATED, FAILED, or as COMPLETED (i.e., market approved). Each simulation step represents one month. At a given point in time, a project in the simulation can be in any of the 9 different states outlined, along with allowed transitions, in Figure 2.
Project states and transitions used in the simulation

As illustrated in Figure 2, in the first step of the simulation, all projects start in a FUNDRAISING state because they are assumed to enter pre-clinical through small- or medium-sized enterprises (SMEs), i.e., developers with “limited” funds for R&D. The simulation allows exits to, or partnerships with, a large pharmaceutical organization (modeled as an agent with “infinite” funds) after the pre-clinical stage. Securing funds for a project can be achieved through four options: public grants, Venture Capital (VC) investments, partnerships and exits. First, a project may receive any grant it is eligible for. If the project receives a grant it will transition to DECIDING, since it now has some funds to initiate development. If the project does not receive a grant the project then pursues one of the remaining three options (VC investment, partnership or exit). The choice of which of these three alternatives is pursued during the time step is selected at random.
If, after FUNDRAISING, the project does not have sufficient funds to perform the next step (namely one month's worth of development), it transitions to HIBERNATING (Figure 2).
If the project secures a VC investment, we assume that it will receive the capital necessary to complete the impending R&D phase in full, but no more. Thus, the project transitions to DECIDING “without infinite funds”. If the project secures a partnership or an exit with a large organization, we assume that it will receive the capital necessary to develop the project to completion. Thus, it transitions to DECIDING “with infinite funds” (Figure 2).
At the DECIDING state, the ENPV of the project is calculated from the perspective of a hypothetical actor representing the composite of all actors involved in the project at that point. If the calculated ENPV is at or above the threshold of the composite actor, then the project transitions into DEVELOPING, if not the project is TERMINATED.
The project will remain in DEVELOPING as long as it has the necessary funds and has not yet reached a decision-point. A project that has funding and reaches a decision-point transitions to DECIDING; whereas a project without the necessary funds transitions to FUNDRAISING.
The agents and decision makers in this simulation are developers and investors in antibiotics R&D (SMEs, VCs, and large pharmaceutical organizations) characterized by different attributes. We assume that large organizations have infinite funds but employ ENPV based valuation to direct activation of their funds, meaning that opportunity cost is taken into account. Conversely, we assume that SMEs constantly need to acquire capital (at costs ranging from 5% to 30%) to progress development. The cost of capital reflects the discount rates applied by VCs when evaluating projects at different R&D phases.
Agents thus have different discount rates that will be used in their ENPV based valuations. VCs, SMEs, and large pharmaceutical organizations employ different discount rates, but do not vary their discount rates across phases. The discount rate for large organizations is in the range of 8% to 13%, while for SMEs a combination of the discount rate of the involved VCs, if any, is applied.
Lastly, agents employ different profit thresholds that a project must meet if the project is to continue. These thresholds are manifested in the profit beyond a positive ENPV required by the agent. VCs and SMEs have no threshold which means we assume that they are only sensitive to relative and not absolute return on investment, while large pharmaceutical organizations employ thresholds between $50Million and $500Million.
Antibiotics projects progress by moving through a number of sequential phases typical for all pharmaceuticals: Pre-Clinical, Phase1, Phase2, Phase3 Clinical Trials, and Launch Activities (i.e. regulatory review, approval, and plant construction).9 These phases differ in terms of their duration, their cost, and the probability of projects successfully completing the phase. Upon creation of a project, the simulator draws random samples for all these parameters from triangularly distributed data based on Sertkaya et al.10 Data on expected net revenues (i.e., global sales minus costs) were also derived from Sertkaya et al. but corrected for inflation into 2017 USD. The global market includes the U.S. data in Sertkaya et al. along with approximations for the rest of the world with adjustments for the time required for market penetration outside the U.S. We made two further adjustments to the data from Sertkaya et al.: (1) we put the lowest possible revenues at 0, since some antibiotics lack a market at the moment of approval, and (2) we lowered the maximum peak year net revenues from $1540Million to $800Million after discussions with specialist advisors in the DRIVE-AB project. We assume that the free market for any antibiotic increases linearly for 10 years (the guaranteed exclusivity period) or (if remaining) to patent expiry.
All our input data was vetted and accordingly modified during several discussions with an expert panel comprised of representatives from pharmaceutical organizations (large and small), public health agencies, and academia. The full set of input data is reported in Table 1.
The list of parameters and values used as input data in the simulation

Results
We present the results of our simulation while discussing how a balance can be struck between the amount of public spending in the MER and the corresponding increase in LOMA. We begin with the results of an initial sensitivity analysis that indicates that a few key input parameters explain a heavy portion of the variance in LOMA. Table 2 summarizes the range of LOMA that can be achieved by controlling only a single input parameter at a time along with the size of a MER. The input parameters explored are: global revenues, VC discount rate, large pharmaceutical organizations' investment threshold, and R&D costs.
Variance in Likelihood of Market Approval (LOMA) explained by different input parameters

* ranges expressed in percentage points (pp)
** relative percentage differences from strongest LOMA determinant
A heavy portion of the variance in LOMA is explained by projected net revenues, when compared to the portion explained by the remaining parameters. Figures 3 to 6 plot the effect of controlling each individual parameter (x-axis) on LOMA (y-axis). The different lines (in Figures 3 to 6) represent MERs of different sizes, ranging from $0 to $3Billion.
The effect of projected total global net revenues on the likelihood of market approval, grouped by market entry reward size (left = fully delinked, right = partially delinked)

The effect of venture capitalists discount rate on the likelihood of market approval, grouped by market entry reward size (left = fully delinked, right = partially delinked)

The effect of profitability thresholds of large pharmaceutical organizations on the likelihood of market approval, grouped by market entry reward size (left = fully delinked, right = partially delinked)

The effect of total projected r&d costs on the likelihood of market approval, grouped by market entry reward size (left = fully delinked, right = partially delinked)

A MER can clearly offset the negative effects of low revenues on LOMA. For example, in Figure 3 a MER of $3Billion makes a project with virtually no market achieve a higher LOMA than a project with total global revenues of $3Billion. Thus a MER successfully compensates for the lack of sales revenue. Interestingly, LOMA is higher when the same amount of money is given as a MER than when provided through sales. Compare, for example, the 2.5% LOMA in Figure 3 when no MER is given at projected revenues of $1500Million with the 3.25% LOMA when a MER of the same size is given. In the case of an FD MER, the MER yields a higher LOMA than if the same amount would have been provided by sales. In the case of a PD MER, the MER should always be preferable since it is provided in addition to any market revenues (Figure 3).
Low MERs do not decrease LOMA because they're optional for the simulated agents. This results in pharmaceutical organizations declining a MER if the projected market sales are preferable. Ignoring this option would overstate the effect of a MER.
Another conclusion that can be drawn from Figure 3 is that not all antibiotics suffer from insufficient projected revenues. In the input parameter span explored, at increasing sizes of projected revenues, the $0 MER line in Figure 3 converges towards the size of the highest MER. This indicates that to obtain a satisfactory level of LOMA (e.g., 3%), no MER is needed when revenues are sufficient. For example, when introducing a MER of $1Billion to a project with revenues of $500Million, LOMA is improved from ~0.5% to more than 3% (2.5 percentage points, pp, increase). Compare this to the minor improvement from ~0.5% to ~1.5% (in FD) and ~2.2% (in PD) (i.e., 1pp and 1.7pp increase respectively) when introducing a MER of the same size ($1Billion) to a project with revenues of $1.5Billion. Clearly, the project with revenues of $500Million is in much greater need of the MER and consequently the increase in LOMA is much more significant.
Comparing full to partial delinkage, we observe that, as projected revenues decrease, larger MER sizes are required to achieve the same LOMA. For example, the difference in LOMA between an FD MER and a PD MER of $800Million is ~0.4 pp (~2.7% FD and 3.1% PD) at revenues of $1500Million, while ~0.8pp (~2.2% FD and ~3% PD) at revenues of $500Million. Meaning that PD is ~36% more effective than FD at low revenues but only ~15% at high revenues.
The second strongest determinant of LOMA is the discount rate of VCs. The LOMA delta of this parameter, as reported in Table 2, is ~31% the size of the strongest determinant. This parameter is plotted for MER sizes of $0-$3Billion in Figure 4.
Under high VC discount rates, SMEs struggle to secure the funding necessary to keep pursuing their projects. In our model, SMEs who fail to secure funding will let projects hibernate indefinitely. Therefore, as shown in Figure 4, as the discount rate of VCs increases, LOMA decreases.
Figure 4 shows that LOMA, in our model, starts to heavily decrease over discount rates of approximately 16%. Below that point, LOMA is not heavily influenced by variations in VC's discount rate.
Let us refer to the area to the left of 18% as low discount rates, and the area to the right as high discount rates. For both low and high discount rates, the MER has a potential effect of increasing LOMA that is capped to some particular point on the y axis (3.5% in Figure 4). This “cap” is what causes the effect on LOMA of high MERs to appear negligible. Indeed, this cap occurs when the average antibiotic project is sufficiently stimulated. At high discount rates fewer MER sizes hit the cap (consider e.g. MERs of $1Billion) than at low discount rates. The implication of this is that the effect on LOMA of some MER sizes is greater when combined with higher, rather than lower, discount rates. This means that a MER of $1Billion, for example, might be unnecessarily high and that $800Million would suffice if VC discount rates were low (≤18%).
When comparing full to partial delinkage, we observe that, at high discount rates, larger MER sizes are required to achieve the same LOMA. For example, the difference in LOMA between an FD MER and a PD MER of $800Million is ~0.15pp (~3.2% FD and ~3.35% PD) at an 18% discount rate, while ~0.5pp (~2.25% FD and 2.75% PD) at a 26% discount rate. Meaning that PD is ~22% more effective than FD at high discount rates but only ~5% at low discount rates.
The next determinant of LOMA according to Table 2 is the profitability threshold of large pharmaceutical organizations, with a LOMA delta of ~20% the size of the strongest determinant. This parameter is plotted for MER sizes between $0 and $3Billion in Figure 5.
Clearly, higher profitability thresholds cause LOMA to decrease. Even though simulations with more data points would be needed to determine the exact nature of this correlation, we can already notice that, under the input range considered, Figure 5 does not exhibit intervention size specific “caps” per intervention size as does Figure 4.
When comparing full to partial delinkage, we observe that, as the threshold increases, larger MER sizes are required to achieve the same LOMA. For example, the difference in LOMA between an FD MER and a PD MER of $800Million is ~0.15pp (~3.3% FD and ~3.45% PD) at a $50Million threshold, while ~0.35pp (~2.55% FD and 2.9% PD) at a $500Million threshold. Meaning that PD is ~5% more effective than FD at low thresholds but ~14% at high thresholds.
The weakest determinant of LOMA is the cost of R&D with a delta reported in Table 2 of ~17% the size of the strongest determinant. This parameter is plotted for MER sizes between $0 and $3Billion in Figure 6.
We observe that in the baseline case (i.e., $0 MER) LOMA seems to linearly decrease with increasing R&D costs, indicating that high-cost projects are especially in need of interventions. However, the positive effect of introducing a MER seems to linearly increase with raising costs. Consequently, high MERs successfully “offset” the negative effects of high R&D costs.
As in Figures 4 to 6, we observe that, as R&D costs increase, FD MERs need to be larger than PD MERs to achieve the same LOMA.
We now shift our focus to specific effects of various sizes and styles of MERs on new antibiotic approvals within a subset of the input parameters. This subset is not only realistic but also makes the effect of and the differences between full and partial delinkage MERs stronger.
This scenario now under study is as follows: large pharmaceutical organizations' ENPV thresholds are between $200Million and $500Million; VC discount rates vary between 18% and 30%; and total projected market revenues per antibiotic is less than or equal to $1500Million globally.
In such a scenario, an FD MER exhibits positive effects over the status quo (LOMA~0.75%). Specifically, the introduction of $800Million MERs doubles the LOMA, from ~0.75% to 1.5% (light boxes in Figure 7).
Changes in likelihood of market approval under different sizes of fully delinked (light) and partially delinked (dark) market entry rewards. Upper outliers truncated.

Increasing the size of the FD MERs steadily increases the LOMA up to $1500-1750Million, where this effect reaches a plateau at ~3% LOMA.
A PD MER starts increasing LOMA earlier than the FD MER, already at $200Million. At a $600Million PD MER the number of new market approvals more than doubles (~1.75% LOMA) compared to the status quo. Market approvals then become more frequent as the PD MER size increase up to a plateau of ~3% LOMA, reached at around $1250Million.
The plateau, shown in Figure 7, indicates that almost all antibiotics, at and beyond this point, are profitable enough to avoid financial termination. Implying that the marginal returns in terms of additional market approvals per additional MER investments might be very low. At this point, any remaining variations of LOMA must be explained by other parameters (e.g., technical probability of success).
In general, Figure 7 suggests that the effect of MERs on LOMA follows a sigmoid curve, and thus suffer from diminishing returns. To exemplify, the marginal improvement in LOMA when moving from a MER of $1000Million to $2000Million is on average significantly lower than when moving from $0 to $1000Million (Figure 7).
The optimal incremental gain of new market approvals (Figure 7) is obtained at $1000Million for FD (from ~1.5% to ~2.7%) and at $800Million for PD (from ~1.75% to ~2.6% LOMA).
Discussion
Our study shows that (1) market revenues have a greater impact than R&D costs on LOMA, and (2) the marginal returns of increased MER sizes exhibit diminishing returns (especially at larger projected net revenues). Therefore, we suggest that (a) interventions improving revenues may be more efficient than those reducing costs, and (b) the balance between cost efficiency (i.e., maximum number of market approvals per dollar) and certainty (i.e., maximum likelihood of market approval per antibiotic) needs to be explored.
A key risk when aiming for certainty is overspending, which entails incentivizing projects that either do not need MERs or whose LOMA is not greatly improved by them. A key risk when aiming for efficiency is under delivering (i.e., not providing sufficient support), which entails failing to incentivize projects that might be of high public health relevance.
At the heart of the question of efficiency versus certainty is the pre-clinical entry rate (or even the discovery rate). If the rate is high then policymakers ought to focus on efficiency. If it is low, then policymakers ought to focus on certainty, because public health cannot afford many terminations before the consequences of antibiotic resistance become too grave. Additionally, when drawing important conclusions on the size and quality of an intervention it may be unproductive and error-prone to merely focus on the average numbers of approved antibiotics, given the heavy variance exhibited in Figure 7. Indeed, it is likely that any given sampled reality (i.e., antibiotic in development) is not close to the mean. If reality happens to be one of the lower outliers, then there is a considerable risk that, in trying to optimize the size of an intervention we miss to stimulate potentially life-saving antibiotics.
Our study shows that (1) market revenues have a greater impact than R&D costs on LOMA, and (2) the marginal returns of increased MER sizes exhibit diminishing returns (especially at larger projected net revenues). Therefore, we suggest that (a) interventions improving revenues may be more efficient than those reducing costs, and (b) the balance between cost efficiency (i.e., maximum number of market approvals per dollar) and certainty (i.e., maximum likelihood of market approval per antibiotic) needs to be explored.
Our simulation indicates that projected revenues seem to be a much stronger determinant of LOMA than R&D costs. This finding is particularly interesting since R&D costs have been emphasized as one of the major hurdles of antibiotics R&D.Reference Katz, Mueller, Polyakov and Weinstock11
We have showed that providing MERs to antibiotics with high revenues, on average, yield a weak marginal improvement of LOMA. If aiming for efficiency, perhaps intervention eligibility should be determined not only on the basis of scientific target product profile (TPP) but also of an antibiotics project's inherent profitability. However, as the future profitability of a new antibiotic may be difficult to forecast, a more viable approach could consist of partial repayments of a PD MER if market sales exceed a given threshold.
In the selected scenario (Figure 7), LOMA can be improved by introducing an FD MER between ~$600Million and ~$1500Million, or by introducing a PD MER between ~$200Million and ~$1250Million. However, as exhibited by the sensitivity analysis, the improvement in LOMA and the difference between FD and PD is higher when conditions are harsh (i.e., high thresholds, discount rates, costs, and low revenues). Thus, the search for an optimal policy intervention is tightly coupled to the characteristics of the target antibiotics and the landscape in which they operate.
The results, in terms of new approved antibiotics as presented in this paper, assumes an entry of new antibiotics in the pre-clinical pipeline. Assessing such a rate of entry was beyond the purpose but should be the object of further research in order to make the trade-off in efficiency-certainty more directly applicable.
Regarding limitations, antibiotics enter our simulation in a stage of pre-clinical development under SME ownership. Consequently, we are neglecting discovery by large organizations and do not explore the effects of interventions on basic research. Further, the reliance of our simulator on ENPV formulas as the key decision rule in R&D is certainly a limitation, as it neglects other decision logics such as project portfolio management, or non-financial logics such as corporate social responsibility. These alternative decision and discovery rules require further research before they can be modeled. We believe that our choice of a rather broad span for all major input parameters (see Table 1) counterbalances the lack of detailed data and helps represent the heterogeneity of projects and developers in the antibiotic field. However, further research on how the various parameters are related to each other (e.g., R&D costs with success probabilities or market revenues) would help to improve the simulation. Finally, while the sensitivity analysis of the effect of different parameters on LOMA has yielded interesting results, the heavy variance exhibited in Figure 7 suggests the need for future work to analyze the same simulation output using correlation techniques.
In conclusion, we confirm the great value that pull-based interventions have in stimulating antibiotics R&D. However, heavy heterogeneity, diminishing returns, and the importance of entry rates suggest that further research is needed before specific sizes and types of market entry rewards can be recommended.
Acknowledgments
The research leading to these results has received support from the Innovative Medicines Initiative Joint Undertaking under grant agreement n°115618 [Driving re-investment in R&D and responsible antibiotic use – DRIVE-AB – www.drive-ab.eu], resources of which are composed of financial contribution from the European Union's Seventh Framework Programme (FP7/2007-2013) and EFPIA companies' in kind contribution. This work does not necessarily represent the view of all DRIVE-AB partners.









