betaFeedback

WorksAgricultural and Biological Sciences

Optimising Recruitment and Sample Allocation in Multi-Site Rare Disease Studies

Abstract

Clinical research in rare diseases is challenged by the limited availability and broad geographical distribution of eligible patients. This necessitates multi-centre collaboration and strategic selection of analytical laboratories based on capacity, cost, and capability. We present a mathematical framework based on cardinality optimisation to support key operational decisions in study design, balancing clinical relevance, feasibility, budgetary constraints, and logistical factors. The framework was implemented in the Reconstruction and Computational Modelling for Inherited Metabolic Diseases (Recon4IMD) project, aiming to recruit 1,900 diagnosed patients across 24 healthcare centres and eight metabolomic laboratories. Compared to both proportional heuristic and greedy marginal allocation baselines, the proposed optimisation model consistently achieved higher recruitment yields in simulated multi-centre scenarios and reduced metabolomic analysis costs by approximately 30% under unconstrained conditions. Notably, both the heuristic and greedy allocation baselines became infeasible under progressively tighter laboratory capacity constraints, whereas the optimised model maintained feasibility and cost control. These findings highlight the value of integrated optimisation strategies for improving recruitment, platform selection, and sample allocation in complex, resource-limited research settings.

affiliation: aff-34 | Maastricht University | Laboratory of Clinical Genetics | |

Introduction

Inherited metabolic disorders (IMDs) are rare genetic diseases that pose significant recruitment challenges due to their geographic dispersion and the need for specialised laboratories for diagnosis and monitoring. Multi-centre clinical studies have become indispensable in medical research, enabling faster recruitment and more generalisable results than single-site studies [1]. In rare diseases, no single hospital can enrol a sufficiently large cohort for meaningful results, making collaboration across multiple sites essential for achieving adequate sample sizes [2]. However, these advantages come with operational challenges, particularly in patient recruitment and sample allocation.

Recruitment in rare disease studies often remains slow, with each site typically contributing only a few patients. Ensuring consistent patient selection and diagnostic criteria across sites adds complexity, especially when expert consensus is required. Designing a multi-centre study thus involves balancing the inclusion of many centres against slow, diffuse recruitment [3]. Laboratory selection also matters: inter-site assay variability is common [4], prompting many studies to centralise testing or harmonise protocols to improve data quality. Previous research has produced piecemeal solutions for these issues, including better sample size formulas and calls for centralised laboratories [5].

Multi-site studies are resource-intensive, with regulatory, ethical, and logistical requirements often limiting expansion [6]. Continuous management of recruitment strategies and data across sites is needed to keep studies on track [7]. Together, these challenges make multi-centre study planning a complex, multi-factorial decision problem.

Mathematical optimisation offers a powerful framework for addressing such complexity, formulating problems as the minimisation or maximisation of objectives under constraints. Within this, cardinality optimisation focuses on selecting an optimal subset of elements, minimising or maximising the number of sites, labs, or allocations. This approach, which enforces parsimony in solutions, has been effective in computational biology, such as in metabolic modelling [8], and shows promise for high-dimensional decision spaces.

Despite these advances, cardinality optimisation has not yet been integrated into operational decision-making for clinical research. While computational frameworks exist for simulating studies [9], there is a lack of applications addressing practical multi-centre questions, such as site selection, lab allocation, or logistical minimisation in a unified way. Clinical study planning still relies on heuristics, isolated analyses, and scenario iteration, rather than a comprehensive algorithmic framework. This gap presents an opportunity to improve multi-centre study design by integrating cardinality optimisation.

Traditional planning approaches often decompose study design into sequential decisions, such as defining sample-size targets, estimating recruitment capacity, and then addressing operational or logistical requirements separately [1, 6, 7]. This separation is practical but limited when decisions are interdependent. In multi-site rare disease studies, recruitment at one healthcare provider changes disease-specific cohort totals, downstream biomarker demand, analytical platform use, laboratory capacity, cost, and sample shipment patterns across the network. Laboratory allocation problems also introduce cost and capacity constraints that are difficult to address through recruitment planning alone [10]. Greedy strategies tend to prioritise local availability or immediate marginal gain, without jointly accounting for cross-disease dependencies, shared capacity constraints, clinical interest, and logistical trade-offs [11, 12]. These limitations motivate an integrated optimisation framework in which recruitment, biomarker demand, platform allocation, cost, and logistics are represented within a single constrained model.

The optimisation framework integrates recruitment planning and sample allocation within a single mathematical model. Disease-specific patient availability across healthcare providers is combined with recruitment capacity, cohort-size targets, biomarker requirements, analytical platform capabilities, laboratory throughput, analysis costs, logistical constraints, and hospital-specific clinical interests. Recruitment decisions define disease-specific cohort sizes and determine biomarker measurement demand, which is then linked to allocation across targeted and global metabolomic platforms. The unified objective combines cardinality terms, recruitment yield, logistical deviation, cost, and clinical-interest alignment to identify feasible and efficient multi-site recruitment and sample allocation strategies. Using the Reconstruction and Computational Modelling for Inherited Metabolic Diseases (Recon4IMD) project as a case study, we show how this strategy supports cost-efficient, logistically feasible, and clinically robust decision-making in complex multi-centre studies. The full mathematical formulation is provided in the Methods section.

Results

The following sections report the application of the proposed mathematical modelling framework to the Recon4IMD project. We first identify which IMDs are feasible for inclusion, based on combined genetic, metabolomic, and recruitment criteria. We then present an overview of patient recruitment and sample allocation across participating healthcare providers and metabolomic platforms. Performance benchmarking compares the cardinality-optimised approach to proportional heuristic and greedy baseline strategies for both recruitment and sample allocation, evaluating recruitment yield, cost efficiency, and feasibility under varying laboratory constraints. Finally, we analyse trade-offs between clinical alignment and logistics, highlighting the flexibility of the optimisation model in balancing competing objectives.

Feasible inherited metabolic diseases in the Recon4IMD project

Three databases, VMH, IEMbase, and ICIMD, were investigated and compared for inclusion in Recon4IMD. The comparison focuses on the overlap of causal IMD genes and known metabolite biomarkers across the datasets. Figure 1a shows a Venn diagram illustrating the overlap of gene sets across the three databases, using gene symbols as the common namespace. The VMH database contains approximately 60% of the IMD causal genes listed in ICIMD, and over 50% of those catalogued in IEMbase. Notably, IEMbase includes over 99% of the causal genes present in ICIMD. All three databases share a core set of 838 genes.

Figure 1. Integration and filtering of inherited metabolic disease data across databases and platforms. This figure illustrates the harmonisation of genetic and metabolic data from multiple sources to identify suitability for metabolic modelling. a) Overlap of gene sets among the VMH, ICIMD, and IEMbase databases using gene symbols as the common namespace. b) Overlap of metabolite sets among the VMH, IEMbase metabolite biomarkers, and targeted and global metabolomic platforms. Comparison is performed using MetaNetX identifiers. c) Stepwise filtering of Inborn Errors of Metabolism (IMDs) from the IEMbase database. The filtering process evaluates data suitability for metabolic modelling and feasibility for patient recruitment based on gene and biomarker availability in VMH and the presence of measurable metabolite biomarkers.

Similarly, Figure 1b compares metabolite sets, including those in VMH, metabolites measurable via targeted and global platforms, and IMD-specific biomarker metabolites listed in IEMbase. A subset of 190 intersecting metabolites represents those that are included in the genome-scale models, are biomarkers for IMDs, and are measurable through existing targeted and global metabolomic platforms within the Recon4IMD project.

By requiring that an IMD has its causal gene and at least one metabolite biomarker represented in the VMH database, and that its biomarker metabolites are measurable using available analytical platforms, we identified the set of IMDs feasible for inclusion in the Recon4IMD project. Figure 1c illustrates this stepwise filtering process. The outermost circle (n=1726n = 1726) includes all IMDs in IEMbase that are annotated with at least one causal gene. The next subset (n=897n = 897) includes IMDs with at least one causal gene present in the VMH database. The third circle (n=365n = 365) further restricts the set to IMDs with at least one measurable metabolite biomarker.

In total, 365 IMDs were identified as suitable for modelling based on the presence of at least one causal gene in the VMH database and at least one measurable metabolite biomarker. However, recruitment feasibility was an additional constraint: not all of these IMDs had sufficient numbers of diagnosed patients across the 24 participating healthcare providers. By incorporating recruitment data and clinical insights, a final subset of 131 IMDs was selected, representing diseases with modellable genes and metabolites, with sufficient patient numbers to enable feasible recruitment within the context of the Recon4IMD project, while excluding entities that were merely biochemical traits without clinical correlates (i.e. ’non-diseases’). These results reflect the database contents at the time the work was conducted and may evolve as VMH, IEMbase, and ICIMD are updated.

Recruitment and participation rate

During the first four months of recruitment at Heidelberg University Hospital, we recorded patient consent outcomes. We also documented reasons for failed sample collection. Of 234 patients approached, 113 consented to provide blood and/or urine samples, giving a participation rate of 48.3%, which was close to the initial assumption of 50%. Figure 2 shows the reasons for non-inclusion. Among the 121 non-included patients, the most frequent reasons were missing signature (30%), reflecting the legal requirement in Germany for both parents to provide consent for minors; language barrier (20%), where informed consent could not be obtained due to communication difficulties; and clinical preference (20%), when treating physicians decided not to approach an eligible patient in order to avoid overburdening or to minimise blood draws in children. These top three reasons accounted for 70% of non-inclusions. The remaining cases were attributed to did not attend (15%), when patients failed to appear for scheduled visits; missing genetic confirmation (10%), referring to patients without confirmed molecular diagnosis such as legacy PA/MSUD cases; and declined participation (5%), when the patient or family actively refused. Grouped by theme, structural or regulatory factors represented 40% (missing signature, missing genetic confirmation), clinical factors 20% (clinical preference), logistical or communication barriers 35% (language barrier, did not attend), and patient preference 5% (declined). Notably, the requirement for consent from both parents applies specifically to paediatric participants in Germany and may differ in other countries.

Figure 2. Reasons for non-inclusion in the study. Bars show the proportion of assessed participants who were not enrolled, ordered from most to least frequent: one-parent consent, language barrier, too unwell or no planned blood draw, did not attend, no genetic test in legacy PA/MSUD cases, and declined participation.

Visualising IMD coverage and sample allocation across the network

To visualise how patient recruitment and sample allocation unfold across disease categories, sites, and platforms under the optimised strategy, we constructed a Sankey diagram. Figure 3 reveals the interplay between ICIMD disease groups, the 24 healthcare providers (HCPs), and the designated analytical platforms. It highlights how disease complexity is managed through coordinated platform use, while balancing biomarker coverage, laboratory constraints, and logistics.

Optimised sample allocation from ICIMD disease groups to healthcare providers and analytical platforms

Figure 3. Optimised sample allocation from ICIMD disease groups to healthcare providers (HCPs) and analytical platforms in the Recon4IMD project. This Sankey diagram visualises the flow of patient samples from inherited metabolic disorders, categorised by ICIMD disease groups (left), through 24 healthcare providers (centre), to analytical platforms (right). Each flow represents the number of patients recruited and assigned for omics analysis under the cardinality-optimised model. Samples are directed to either global platforms (purple) or targeted platforms (blue), based on biomarker coverage, laboratory capacity, and logistical constraints. Flow colours correspond to ICIMD groups, highlighting how disease categories are distributed across sites and platforms. The diagram illustrates the integrated recruitment and sample allocation strategy designed to maximise diagnostic yield while minimising logistical burden.

Benchmarking results

To benchmark recruitment performance, we simulated 24 scenarios with increasing numbers of HCPs, ranging from single-centre to full-network participation. In each scenario, the cardinality-optimised model was compared against two baseline recruitment strategies: the proportional heuristic approach (Algorithm S1) and the greedy recruitment baseline (Algorithm S2). All methods were applied to the same set of hospitals in each scenario, and the total number of recruited patients was recorded. This comparison focused exclusively on recruitment performance and did not incorporate downstream allocation or cost considerations.

The results, shown in Figure 4, indicate that while all approaches performed similarly in single or two-centre configurations, differences emerged as network size increased. The proportional heuristic recruitment strategy exhibited limited ability to exploit cross-site complementarities, leading to reduced recruitment totals in larger networks. The greedy recruitment baseline demonstrated improved coordination relative to the proportional heuristic; however, it remained locally driven and did not achieve the same recruitment levels as the cardinality-optimised model. As network complexity increased, the performance gap between the optimisation model and both baseline strategies widened, reflecting the growing importance of global coordination across multiple diseases and hospitals.

Optimisation gain vs heuristic (HCP engagement)

Figure 4. Optimisation gain vs heuristic (HCP engagement). Blue circles represent the proportional heuristic recruitment strategy (Algorithm S1), red triangles denote the global greedy recruitment baseline (Algorithm S2), and yellow squares correspond to the cardinality-optimised model. The shaded green region indicates the recruitment gain achieved by the optimisation model relative to the global greedy recruitment baseline.

Under unconstrained laboratory capacity conditions, the recruitment and allocation model achieved approximately a 30% reduction in total metabolomic analysis cost compared to the heuristic disease-wise allocation strategy (Algorithm S3). The global greedy marginal allocation baseline (Algorithm S4) reduced costs relative to the heuristic approach but remained more expensive than the cardinality-optimised solution. Cost savings were observed for both targeted and global platforms. Figure 5 presents the comparative cost breakdown across all three allocation strategies.

Figure 5. Metabolomic analysis cost comparison for 843 patients. Blue bars represent the heuristic disease-wise allocation strategy (Algorithm S3), orange bars denote the global greedy marginal allocation baseline (Algorithm S4), and green bars correspond to the cardinality-optimised recruitment and allocation model. Costs are reported in thousand euros and are decomposed into targeted platform costs, global platform costs, and total analysis cost.

Figure 6 illustrates the feasibility landscape across varying targeted and global platform capacity constraints. Each grid cell represents a specific combination of platform throughput limits, classified according to whether the optimisation model, the heuristic allocation strategy (Algorithm S3), and the greedy allocation baseline (Algorithm S4) are feasible. Green regions indicate configurations where all approaches identify valid allocations, yellow regions denote cases where only the cardinality-optimised model remains feasible, and red regions correspond to scenarios where no method can satisfy the coverage requirements. Notably, the heuristic and greedy baselines became infeasible at identical capacity thresholds, whereas the optimisation model remained feasible over a strictly larger region of the parameter space. Across all tested capacity levels, there were no instances in which either baseline method was feasible while the optimisation model failed, highlighting the robustness of the proposed formulation under progressively tighter operational constraints.

Feasibility comparison of sample allocation strategies

Figure 6. Feasibility comparison of sample allocation strategies. The x-axis and y-axis represent the maximum number of samples that can be analysed by the targeted and global metabolomic platforms, respectively. Each cell corresponds to a tested combination of platform capacity limits. Green regions indicate configurations where the optimisation model, the heuristic allocation strategy (Algorithm S3), and the greedy allocation baseline (Algorithm S4) all return feasible solutions. Yellow regions denote cases where only the cardinality-optimised model remains feasible, while red regions correspond to scenarios in which none of the methods can satisfy the coverage requirements.

Trade-off between clinical interest and logistic overhead

Figure 7 illustrates the trade-off between clinical alignment and logistics cost. As the clinical interest weight λ\lambda increases, the recruitment pattern increasingly reflects each hospital’s interest in specific IMDs. However, this improved alignment comes at the cost of reduced local processing and increased inter-hospital sample shipments. The resulting curve traces out a meaningful Pareto frontier, confirming the model’s ability to capture the inherent tension between clinical relevance and logistical efficiency.

Figure 7. Clinical interest vs logistic overhead. Pareto frontier illustrating the trade-off between two competing objectives: the number of recruited patients aligned with hospital-specific clinical interests (left axis, circles) and logistic overhead which is a summation of logistic deviation variables (right axis, squares). Each point represents the solution at a different value of the clinical interest coefficient (λ\lambda), ranging from 1 to 1000.

The curve shows a clear pattern of diminishing returns. Initial increases in λ\lambda produce marked gains in clinical-interest alignment, whereas higher values produce smaller additional gains while logistics overhead continues to rise. Using the Pareto-knee criterion described in the Methods, λ=70\lambda= 70 was selected for the primary analyses. This value represents the point at which clinical-interest alignment is incorporated without imposing disproportionate logistics overhead.

Discussion

Despite common concerns about the computational cost of optimisation-based methods [10], our results suggest that, with appropriate vectorisation, sparse reformulation, and solver selection, such models can scale efficiently in large multi-centre study settings. In the Recon4IMD case study, the model remained computationally tractable as recruitment, allocation, capacity, cost, logistics, and clinical-interest components were progressively incorporated. Supporting metrics, including model dimensionality and solving times across model versions, are provided in Supplementary Table S1. More generally, computational burden is expected to increase with the number of diseases, healthcare providers, biomarkers, and analytical platforms, since these determine the size of the decision-variable vector and associated constraint matrices. These considerations are important when adapting the framework to larger rare disease networks.

Beyond efficiency gains, mathematical optimisation frameworks improve fairness and transparency in operational decisions. In multi-site studies like Recon4IMD, decisions regarding patient recruitment and sample allocation are effectively a zero-sum game, where resources such as patient access or analytical capacity at one site are allocated at the expense of another. This creates potential for competing institutional incentives and places considerable pressure and responsibility on the project’s coordination unit to make equitable decisions under competing demands. Moreover, budget holders require clear justification for how resources are distributed, particularly when public or pooled funding is involved. An optimisation-based approach addresses these concerns by offering a formally defined, reproducible, and data-driven framework for decision-making. It enables the simultaneous consideration of multiple factors (such as recruitment feasibility, analytical throughput, cost, logistics, and clinical interest) that would be difficult for any individual or committee to balance manually. In doing so, it not only improves operational robustness but also promotes stakeholder trust by providing transparent and justifiable allocation outcomes.

In multi-site clinical studies, optimisation-driven allocation must also consider potential sources of systematic bias. Recruitment patterns may unintentionally overrepresent certain geographic regions, demographic groups, or clinical subpopulations, potentially conflicting with diversity, equity, and inclusion objectives or regulatory expectations. Likewise, heterogeneity in site-level performance, such as protocol deviation rates, screen failure patterns, or competing trial involvement, may introduce differences in data quality, and correlations between disease distribution and laboratory assignment could confound biological signals with platform effects. The present study did not include individual-level demographic or site-quality data, and therefore, such fairness-aware or bias-mitigation constraints were not incorporated into the optimisation model. However, the linear constraint-based structure of the framework readily permits extension through additional representation constraints, site-level weighting factors, or platform-balancing requirements where such data are available. For example, future implementations should consider lower and upper bounds on recruitment by age group, sex, geographic region, or under-represented subgroup, platform-balancing constraints to avoid confounding biological subgroups with laboratory assignment, and site-quality weights based on protocol-deviation or screen-failure rates. Incorporating these elements would enable bias-aware optimisation while preserving transparency and formal rigour in allocation decisions.

While the optimisation-based framework offers considerable benefits in terms of efficiency, fairness, and transparency, it is important to acknowledge that model feasibility is not guaranteed under all circumstances. Specifically, if the constraints imposed on the system are mutually incompatible, the optimisation problem will have no solution. For example, infeasibility can arise if the disease-specific cohort size requirements, healthcare provider recruitment capacities, and site-specific patient availability do not admit a distribution of patients that simultaneously satisfies all constraints. This could occur if certain diseases require more patients than are available across the network, or if site-specific capacities are too restrictive to permit allocation that meets cohort targets. In practice, recognising and diagnosing such infeasibility is a key advantage of mathematical optimisation. Empirically, our feasibility landscape analysis demonstrated that the optimisation model remained feasible over a strictly larger region of the operational parameter space than either baseline strategy. Importantly, within the tested platform-capacity configurations and under identical cohort definitions, no scenario was observed in which either baseline method achieved feasibility while the optimisation model failed. Accordingly, model outputs should always be interpreted in light of the input data and constraints, and users are encouraged to perform sensitivity analyses to identify constraint configurations that preserve feasibility. In addition to structural infeasibility arising from model constraints, practical implementation also faces several real-world barriers related to patient availability and participation. The integrality checks confirmed that all reported recruitment and allocation quantities were integer-valued within solver tolerance and therefore interpretable as patient and sample counts.

However, beyond strict feasibility considerations, the stability of optimal allocations under changing assumptions also warrants attention. Because the framework incorporates cardinality objectives, small perturbations in input parameters, such as updated patient availability, capacity limits, or biomarker mappings, may lead to discrete structural changes in the optimal solution. This sensitivity is characteristic of non-convex optimisation and may result in substantially different recruitment or platform assignment patterns when assumptions are revised. In operational settings, re-optimisation should therefore be performed with caution. Rather than recalculating allocations from scratch, previously executed or contractually committed decisions may be fixed as additional constraints, preserving continuity of ongoing recruitment or laboratory processing. The framework readily accommodates such constraints by restricting selected decision variables during iterative re-optimisation. In this way, adaptive planning can be achieved while maintaining operational stability. Although extensions based on robust or stochastic optimisation could explicitly model parameter uncertainty, such formulations would substantially increase computational complexity and were beyond the scope of the present study.

In addition to these structural considerations, real-world implementation faces further barriers, particularly in rare disease contexts, where patient availability, consent logistics, and participation rates may fall short of expectations. To address this, we adjusted our model inputs to reflect an estimated number of realistically recruitable patients. This enhances the practical relevance of the optimisation outcomes by aligning them with expected recruitment attrition rates observed in real-world rare disease studies, thereby improving the reliability of projected cohort feasibility.

Mathematical optimisation provides a structured basis for planning, but its practical implementation in clinical research remains inherently collaborative. In real-world settings, certain decisions, such as adjusting to unforeseen staffing shortages, regulatory delays, or evolving clinical priorities, may fall outside the scope of even the most comprehensive models. Rather than replacing human oversight, the optimisation framework is best understood as a decision support tool that streamlines and formalises resource allocFation, while still requiring refinement through expert judgement. This human-in-the-loop approach enables research coordinators to adapt optimised plans to live operational conditions, ensuring both robustness and contextual appropriateness. By combining algorithmic rigour with clinical insight, this strategy fosters more resilient and resource-efficient study coordination. In the present analysis, robustness was assessed through sensitivity analyses of HCP engagement, laboratory capacity limits, and the clinical-interest weighting parameter. Future efforts should continue to integrate dynamic operational data to further align optimisation outputs with the evolving realities of clinical research.

This study demonstrates the value of a unified optimisation framework for patient recruitment and sample allocation in complex, multi-site rare disease studies. Applied to the Recon4IMD project, the approach integrated cardinality optimisation, cost and logistics modelling, and clinical alignment, enabling robust and efficient operational planning across a European research network. Compared to proportional heuristic and greedy baseline strategies, the optimisation model consistently achieved higher recruitment yields, remained feasible under tighter capacity constraints, and reduced overall metabolomic analysis costs. The flexibility to incorporate multiple objectives and operational constraints supports transparent and data-driven decision-making in clinical research, offering a scalable solution for future large-scale, resource-limited studies. As database resources, omics technologies, and cross-border research infrastructures continue to evolve, optimisation-based planning frameworks will be essential for coordinating increasingly complex, multi-objective biomedical collaborations under real-world resource constraints.

Methods

This section outlines the mathematical framework used to model clinical recruitment and metabolomic sample allocation. Central to our approach is the use of cardinality optimisation, which enables the selection of a minimal set of recruiting centres or resources while satisfying estimated patient counts, recruitment capacity, and cohort targets. We begin by formulating a general recruitment plan as a cardinality optimisation problem, based on estimated patient count data, site-specific recruitment capacity constraints, and disease-specific cohort recruitment targets. Subsequent subsections expand the optimisation model to incorporate sample allocation to analytical laboratories, cost-efficient platform selection for omics analysis, logistical considerations, and alignment with the clinical interests of healthcare providers. Finally, the general framework is applied to the Recon4IMD project, with a detailed explanation of how feasible IMDs were selected for investigation and how the optimisation model was adapted to this case. The full mathematical formulation, including the vectorised implementation, is provided in the Supplementary sections ’Detailed Formulation of Sample Allocation Optimisation in Recon4IMD’ and ’Transforming optimisation formulas for computational implementation’.

Notation convention: Input data matrices are denoted using bold capital letters (e.g. D\mathbf{D} for the estimated patient count matrix). Variable matrices are represented by capital italic letters (e.g. X,YX,Y), while scalar and vector variables are shown as lowercase italic and Greek letters (e.g. yy, β\beta). While the same notation is used throughout the general mathematical formulation and the Recon4IMD case study, specific variables may correspond to particular datasets or constraints in the applied context. A full summary of notation is provided in Supplementary Table S3.

Clinical recruitment optimisation

To enable effective patient recruitment across hospitals, the optimisation framework begins with detailed hospital-level estimated patient count data for each disease. These data represent the number of potentially accessible patients at each site and serve as the empirical basis for the recruitment model. They are used as input in the mathematical formulation. The model links the estimated patient counts to recruitment allocation decisions, such as how many patients to recruit per disease at each hospital, while accounting for site-specific capacity and feasibility constraints.

Let D∈Nm×n\mathbf{D} \in\mathbb{N}^{m \times n} be an estimated patient count matrix where Di,j\mathbf{D}_{i,j} is the estimated number of patients of disease ii at healthcare provider (HCP) jj. Let X∈Rm×n\mathbf{X} \in\mathbb{R}^{m \times n} be a recruitment fraction matrix where Xi,j∈[0,1]X_{i,j} \in[0,1] is the fraction of patients of disease ii recruited at HCP jj. Let Y∈Rm×n\mathbf{Y} \in\mathbb{R}^{m \times n} be a recruitment number matrix where Yi,j∈RY_{i,j} \in\mathbb{R} represents the planned number of patients with disease ii recruited at HCP jj. Then we have

D∘X=Y,(1)\mathbf{D} \circ\mathbf{X} = \mathbf{Y}, \tag*{(1)}

where ∘\circ denotes the Hadamard product, that is, Di,j×Xi,j=Yi,j\mathbf{D}_{i,j} \times X_{i,j} = Y_{i,j}. The recruitment numbers are modelled as continuous decision variables representing planned allocation levels. Although patient counts are inherently integer-valued, the optimisation problem is formulated as a continuous relaxation to improve computational tractability within the integrated recruitment and allocation framework. Because patient and sample counts are discrete, final recruitment and allocation outputs were checked for integrality by comparing each value with its nearest integer within solver tolerance before reporting the solution. No non-integer recruitment or allocation values were observed in the reported analyses.

Minimising the number of active disease–HCP recruitment assignments, thereby encouraging each disease to be recruited at fewer sites and reducing fragmentation of recruitment efforts, while simultaneously maximising the number of patients recruited per disease, can be formulated as the following cardinality optimisation problem:

min⁡X,Y∥X∥0−1Ty,s.t.D∘X−Y=0,Y1−y=0,YT1−z=0,0≤X≤1,y‾≤y≤y‾,z‾≤z≤z‾.(2)\begin{aligned} \underset{\mathbf{X},\mathbf{Y}}{\min}\quad& \lVert\mathbf{X} \rVert_{0} - \mathbf{1}^{\mathsf{T}}\mathbf{y},\\ \text{s.t.}\quad& \mathbf{D} \circ\mathbf{X} - \mathbf{Y} = 0,\\ & \mathbf{Y}\mathbf{1} - \mathbf{y} = 0,\\ & \mathbf{Y}^{\mathsf{T}}\mathbf{1} - \mathbf{z} = 0,\\ & 0 \leq\mathbf{X} \leq1,\\ & \underline{\mathbf{y}} \leq\mathbf{y} \leq\overline{\mathbf{y}},\\ & \underline{\mathbf{z}} \leq\mathbf{z} \leq\overline{\mathbf{z}}. \tag*{(2)} \end{aligned}

In this formulation, the term ∥X∥0\lVert\mathbf{X} \rVert_{0} counts the number of non-zero disease–HCP recruitment assignments, that is, the number of disease–site pairs (i,j)(i,j) for which recruitment is active. Minimising ∥X∥0\lVert\mathbf{X} \rVert_{0}, therefore, encourages recruitment for each disease to be concentrated across fewer sites, reducing fragmentation of recruitment activities. Operationally, such concentration can reduce administrative and monitoring burden, simplify coordination and contracting across centres, and support more consistent protocol implementation and data collection. From a statistical perspective, limiting fragmentation may mitigate inter-site heterogeneity in clinical practice and measurement procedures, thereby improving comparability of observations across the network.

The vector y∈Rm\mathbf{y} \in\mathbb{R}^{m} represents the planned total number of patients recruited per disease, and z∈Rn\mathbf{z} \in\mathbb{R}^{n} represents the planned total number of patients recruited by each healthcare provider. Here, 1\mathbf{1} denotes a vector of ones with appropriate dimension. The lower and upper bounds y‾\underline{\mathbf{y}} and y‾\overline{\mathbf{y}} constrain the number of recruited patients for each disease, based on each disease cohort. Similarly, z‾\underline{\mathbf{z}} and z‾\overline{\mathbf{z}} specify the recruitment capacity limits for each healthcare provider, reflecting constraints such as available staff, or infrastructure.

Biomarker demand induced by patient recruitment

As patient samples are destined for omics analysis, the patient recruitment and sample allocation strategies must be designed in an integrated manner. Because samples correspond to different inherited metabolic disorders (IMDs), the set of biomarkers to be measured varies by disease. The relationships between diseases and their characteristic biomarkers can be obtained from established domain-specific resources, such as IEMbase [13], as well as from curated clinical literature. These disease-biomarker associations provide the biological foundation for determining measurement requirements within the optimisation framework.

Let B∈{0,1}p×m\mathbf{B} \in\{0,1\}^{p \times m} be a biomarker-association matrix, where Bi,j=1\mathbf{B}_{i,j} = 1 if biomarker ii is associated with disease jj, and Bi,j=0\mathbf{B}_{i,j} = 0 otherwise. Let y∈Rm\mathbf{y} \in\mathbb{R}^{m} denote the planned total number of patients recruited per disease. The resulting biomarker measurement matrix is defined as:

S=Bdiag⁡(y),(3)S = \mathbf{B}\operatorname{diag}(y), \tag*{(3)}

where S∈Rp×mS \in\mathbb{R}^{p \times m} and Si,jS_{i,j} represents the number of times biomarker ii must be measured for disease jj. This formulation ensures that biomarker demand scales proportionally with disease-specific recruitment levels and establishes a direct link between patient recruitment and downstream analytical requirements.

Laboratory allocation and capacity constraints

Designing an efficient omics analysis strategy involves determining a consolidated allocation of analytical platforms required to satisfy the biomarker measurement matrix SS, while respecting laboratory capacity and operational constraints. In addition to coverage considerations, the allocation must reflect the analytical capabilities and throughput limitations of each participating laboratory.

To mathematically model platform capabilities, data are collected from participating laboratories specifying which biomarkers, such as genes, proteins, or metabolites, are detectable by each analytical test panel. These platform-specific biomarker lists are cross-mapped and harmonised across laboratories to ensure consistent identifier usage and enable fair comparison of coverage. Once harmonised, this capability information serves as structured input to the optimisation framework.

Let T∈{0,1}p×qT \in\{0,1\}^{p \times q} be a platform capability matrix where Ti,j=1T_{i,j}=1 if biomarker ii is detectable by analytical platform jj and Ti,j=0T_{i,j}=0 otherwise. Let U∈Rq×mU \in\mathbb{R}^{q \times m} represent platform utilisation where Ui,jU_{i,j} is the number of samples from disease jj assigned to platform ii.

To ensure that all required biomarkers are measured, the allocation must satisfy:

TU≥S.(4)TU \ge S. \tag*{(4)}

The inequality allows biomarkers to be measured more than once where necessary, providing flexibility in platform assignment.

To reduce fragmentation of disease-platform assignments while ensuring full biomarker coverage, we formulate the following cardinality optimisation problem:

min⁡U∥U∥0,s.t.TU≥S,(5)\begin{aligned} \min_{U}\quad& \lVert U\rVert_{0},\\ \text{s.t.}\quad& TU \ge S, \tag*{(5)} \end{aligned}

Because SS is defined in Eq. (3) as a function of the planned recruitment vector yy, any non-zero recruitment implies S≠0S \ne0, and therefore the trivial solution U=0U=0 is infeasible. The term ∥U∥0\lVert U\rVert_{0} counts the number of non-zero disease–platform assignments, that is, the number of disease-specific sample flows routed to analytical platforms. Minimising ∥U∥0\lVert U\rVert_{0} therefore promotes consolidation of analytical workflows, reducing fragmentation of disease-specific measurements across multiple platforms. Operationally, limiting cross-platform splitting can simplify logistics, reduce sample handling complexity, and streamline coordination between laboratories. Thus, ∥U∥0\lVert U\rVert_{0} is interpreted as a sparsity penalty on individual disease–platform assignments.

This objective promotes assignment-level sparsity. Introducing explicit platform-level binary selection variables would lead to a substantially more complex mixed-integer formulation and was therefore avoided to preserve computational tractability. The present formulation balances analytical consolidation with computational tractability while maintaining flexibility for platform reuse across multiple diseases.

To reflect real-world operational limits, laboratory throughput constraints are also introduced. Each analytical platform is assumed to have a finite processing capacity, restricting the total number of patient samples that can be analysed by that platform. To model this, an auxiliary variable is defined. Let u∈Rqu \in\mathbb{R}^{q} be the total number of samples allocated to each analytical platform, where uiu_i represents the total number of samples assigned to analytical platform ii. This auxiliary variable is linked to the platform utilisation matrix through the relation:

u=U1m,(6)u = U\mathbf{1}_{m}, \tag*{(6)}

where 1m\mathbf{1}_{m} denotes a column vector of ones of length mm, corresponding to the number of diseases.

Capacity limits are imposed as:

0≤ui≤uˉi,∀i=1,…,q,(7)0 \le u_i \le\bar{u}_i,\quad\forall i=1,\ldots,q, \tag*{(7)}

where uˉi\bar{u}_i is the maximum number of samples that analytical platform ii is capable of processing. This formulation ensures that biomarker demand induced by recruitment is satisfied, while reducing disease-platform allocation fragmentation and respecting operational capacity constraints.

Cost and logistics modelling

In many multi-centre omics studies, multiple analytical laboratories participate, each offering different platforms, operating under distinct technical and logistical conditions. The cost of omics analysis is therefore, both laboratory-specific and platform-dependent. Additional factors such as the number of samples, batch sizes, and sample types further influence the overall cost of analysis. As with most clinical research projects, budget constraints play a critical role in shaping sample allocation strategies. Consequently, mathematical modelling of sample allocation should incorporate the associated analytical costs to ensure that the selected strategy remains within budget.

Let r∈R1×q\mathbf{r} \in\mathbb{R}^{1 \times q} denote the platform cost vector, where rir_i represents the cost of performing one omics analysis using platform ii. Let β∈R1×m\boldsymbol{\beta} \in\mathbb{R}^{1 \times m} represent the total omics analysis cost per disease, where βj\beta_j corresponds to the overall cost associated with analysing samples from disease jj. The relationship between platform utilisation and disease-specific costs is defined as:

rU=β.(8)\mathbf{r}U = \boldsymbol{\beta}. \tag*{(8)}

This formulation assumes a linear cost model, where total costs scale directly with the number of assigned analyses, and does not account for potential non-linearities such as batch effects, minimum order sizes, or setup overhead.

Incorporation of clinical interest in the recruitment strategy

In addition to the logistical and operational factors previously considered in the recruitment strategy, the clinical interests of each healthcare provider also play an important role. Hospitals with focused expertise in certain diseases are more likely to have experienced staff familiar with the associated diagnostic, treatment, and management requirements. These sites may also be involved in ongoing research or clinical studies related to specific diseases, thereby increasing the likelihood of successful patient recruitment and long-term engagement.

To incorporate site-specific clinical expertise into the recruitment optimisation framework, a linear objective coefficient c∈Rd\mathbf{c} \in\mathbb{R}^{d} is introduced, where dd denotes the total number of decision variables in the vectorised optimisation problem. Specifically, to reflect the clinical interest of each healthcare provider, the coefficient c\mathbf{c} is assigned to each disease-hospital pair (i,j)(i,j) and reflect as

c(i,j)=−λ×C(i,j),(9)c(i,j) = -\lambda\times C(i,j), \tag*{(9)}

where C∈Rm×nC \in\mathbb{R}^{m \times n} is the clinical interest matrix, with C(i,j)=1C(i,j) = 1 if hospital jj is interested in recruiting patients with disease ii, and 0 otherwise. λ\lambda referred to as the clinical interest weight, is a user-defined parameter that reflects the relative importance of clinical interest compared to other objectives in the model. Since the optimisation problem is formulated as a minimisation, the negative sign ensures that clinical interest contributes positively by reducing the overall objective value, thus encouraging the allocation of patients to hospitals aligned with specific disease interests. The parameter λ\lambda, controls the strength of this incentive relative to other objectives in the model.

The selection of λ\lambda was calibrated through a sensitivity analysis over the tested range of clinical-interest weights. For each value of λ\lambda, we recorded two outcomes: the number of recruited patients aligned with declared clinical interest, denoted A(λ)A(\lambda), and the total logistics deviation, denoted L(λ)L(\lambda). Both quantities were min-max normalised to [0,1][0,1]. Since A(λ)A(\lambda) is maximised and L(λ)L(\lambda) is minimised, the selected value of λ\lambda was defined using a Pareto-knee criterion. Specifically, the selected value was the point with the maximum signed perpendicular distance from the chord connecting the two extreme normalised Pareto solutions. This identifies the point at which further improvement in clinical-interest alignment requires a disproportionate increase in logistics deviation. This approach follows established knee-point selection principles for bicriteria Pareto trade-offs [14].

It is important to note that this clinical interest term acts as a soft objective and not a constraint. The optimisation always ensures feasibility based on estimated patient counts and site capacity. When multiple hospitals have sufficient patient populations for a given disease, the model will prioritise assignments to those that have declared clinical interest. If only one among the eligible hospitals expresses interest, that site will be preferred. In cases where no hospital declares interest for a disease, the allocation will still proceed based on feasibility. Thus, the clinical interest term provides a directional preference that guides allocation decisions when alternatives exist.

Modelling sample allocation logistics

In many real-world clinical studies, certain healthcare providers are either directly equipped with in-house omics laboratories or maintain collaborative agreements with nearby analytical facilities. These hybrid sites can both recruit patients and perform omics analyses locally. In such scenarios, it may be inefficient or unnecessary to transport biological samples over long distances when local analysis is feasible. To reflect this consideration, the objective function can be augmented with a penalty for allocating samples to non-local or logistically distant laboratories.

Let Y∈Rm×tY \in\mathbb{R}^{m \times t} denote a subset of the recruitment matrix YY, where tt is the number of hybrid or co-located sites equipped with omics platforms. In an ideal scenario, there would be full alignment between the number of patients recruited at these sites and the corresponding platform utilisation, such that samples collected at a given site are analysed locally. This can be expressed as:

Yi,j=Ui,jT,∀ i=1,…,m, j=1,…,t.(10)Y_{i,j} = U_{i,j}^{T}, \quad\forall\ i = 1,\ldots,m,\ j = 1,\ldots,t. \tag*{(10)}

Deviations from ideal local processing are captured through a logistics deviation matrix Fi,j∈Rm×tF_{i,j} \in\mathbb{R}^{m \times t}, defined as:

Fi,j:=∣Yi,j−Ui,jT∣∀ i=1,…,m, j=1,…,t.(11)F_{i,j} := \left|Y_{i,j} - U_{i,j}^{T}\right| \quad\forall\ i = 1,\ldots,m,\ j = 1,\ldots,t. \tag*{(11)}

Here, Fi,j=0F_{i,j}=0 indicates that samples collected at site jj for disease ii are analysed locally, with no external shipment required. Since such perfect alignment is rarely achievable in practice, this constraint is relaxed:

Fi,j≥∣Yi,j−Ui,jT∣.(12)F_{i,j} \geq\left|Y_{i,j} - U_{i,j}^{T}\right|. \tag*{(12)}

To discourage unnecessary sample shipment and promote local analysis where possible, the model includes a penalty term in the objective function that minimises the total logistical deviation. This is expressed as the one-norm of the deviation matrix:

∥F∥1=∑i=1m∑j=1tFi,j.(13)\lVert F \rVert_{1} = \sum_{i=1}^{m}\sum_{j=1}^{t} F_{i,j}. \tag*{(13)}

Minimising this term supports cost reduction, improves logistical efficiency, and enhances overall robustness of the sample allocation strategy.

Scaling of Objective Components

The integrated objective function combines heterogeneous components, including operational limitations, recruitment totals, logistics deviations, and monetary cost terms. Because these components are expressed in different natural units, their numerical magnitudes may differ substantially. In weighted-sum formulations, components operating at higher orders of magnitude may exert disproportionate influence unless explicitly scaled.

Structural terms such as ∥X∥0\lVert X \rVert_{0}, ∥U∥0\lVert U \rVert_{0}, recruitment totals ∥y∥1\lVert y \rVert_{1}, and logistics deviations ∥F∥1\lVert F \rVert_{1}, ∥W∥1\lVert W \rVert_{1}, are bounded by the number of diseases, sites, platforms, and cohort limits, typically operating at magnitudes on the order of 102−10310^{2} - 10^{3}. In contrast, laboratory cost terms scale with per-sample platform costs and total cohort size, and may operate at magnitudes several orders higher. For instance, assuming an estimated per-sample omics analysis cost of 100 euros, a cohort of approximately 1,000 patients would yield a total laboratory cost ∥β∥1\lVert\beta\rVert_{1} to be on the order of 10510^{5}, which is two orders of magnitude larger than operational components driven by patient or sample counts.

To ensure comparable numerical magnitudes across objective components while preserving interpretability in natural units, a scaling coefficient ρ∈R>0\rho\in\mathbb{R} > 0 is introduced for cost-related l1l_{1} terms. The resulting objective function can be written as:

min⁡ cTφ(v)+∥X∥0−∥y∥1+∥U∥0+∥F∥1+ρ∥β∥1,(14)\min\ c^{T}\varphi(v) + \lVert X \rVert_{0} - \lVert y \rVert_{1} + \lVert U \rVert_{0} + \lVert F \rVert_{1} + \rho\lVert\beta\rVert_{1}, \tag*{(14)}

where ρ\rho controls the numerical scale of monetary cost contributions relative to operational and logistical components. In this formulation, the vector vv contains all decision variables of the model, and φ(v)∈Rd\varphi(v) \in\mathbb{R}^{d} denotes the vectorisation operator that maps these structured variables into a single decision vector compatible with the objective coefficient vector c∈Rdc \in\mathbb{R}^{d}.

The ℓ0\ell_{0}-type terms in the objective, including ∥X∥0\lVert X \rVert_{0} and ∥U∥0\lVert U \rVert_{0}, were handled through an established cardinality optimisation implementation [8, 15]. In this approach, the zero-norm is approximated using non-convex continuous functions, reformulated as a difference-of-convex optimisation problem, and solved approximately through a sequence of convex optimisation subproblems. Gurobi was used as the solver for these subproblems [16]. The implemented model is therefore a cardinality approximation based on a continuous reformulation, rather than an exact mixed-integer or exhaustive enumeration formulation.

Solver diagnostics were recorded for the reported optimisation model variants. Because the zero-norm terms were handled through a continuous cardinality approximation rather than an exact mixed-integer formulation, MIP gaps are not applicable. For each reported model variant, Gurobi returned optimal termination for the implemented optimisation problem, and the final solution satisfied the implemented constraints with zero maximum feasibility residual under the diagnostic checks. These diagnostics, together with the existing model dimensionality and solving-time information, are reported in Supplementary Tables S1 and S2. Accordingly, the reported solutions should be interpreted as feasible approximate cardinality-optimised solutions, rather than certified global optima of the original zero-norm formulation.

The cardinality optimisation framework supports such scaling via user-defined objective coefficients applied to selected entries of the vectorised decision vector φ(v)\varphi(v), allowing controlled adjustment of relative influence without modifying the constraint structure [8]. In the Recon4IMD implementation, we set ρ=10−2\rho= 10^{-2}, chosen to align the numerical magnitude of cost terms with operational components.

Unified optimisation formulation

All previously discussed components, patient recruitment strategy, omics sample distribution, cost minimisation, logistical deviation, and clinical interest alignment, are integrated into a unified optimisation framework. The resulting model balances multiple objectives, including maximising the number of patients recruited across all diseases, minimising the number of recruiting sites and platforms, encouraging local processing, reducing costs, and aligning recruitment with clinical expertise. The complete formulation is illustrated in Figure 8.

Figure 8. Unified optimisation framework for multi-centre clinical study design. This framework integrates clinical recruitment, sample allocation, and logistical constraints into a single mathematical model. a) Patient recruitment strategy: selecting healthcare providers to recruit patients based on disease-specific estimated patient counts at each site, target disease cohorts, and site-specific capacities. b) Sample allocation strategy: assigning biological samples to analytical platforms while ensuring full biomarker coverage and laboratory throughput constraints. c) Logistics and cost optimisation: minimising non-local sample shipments and analytical costs. The objective function incorporates clinical interest through linear objective coefficients and combines cardinality and one-norm terms to balance recruitment, platform usage, logistics, and cost efficiency.

Case study: Recon4IMD implementation

Reconstruction and Computational Modelling for Inherited Metabolic Diseases (Recon4IMD) is a pan-European research initiative aimed at enhancing the diagnosis of IMDs through integrated metabolomic and computational modelling approaches. IMDs comprise over 1,990 rare genetic disorders with diverse and often ambiguous clinical presentations, which complicate diagnosis and delay treatment. Given the limited patient populations per site and the need for advanced analytical workflows, Recon4IMD provides a representative case for optimising recruitment and sample allocation across a distributed network.

To address these challenges, the Recon4IMD project mobilises a network of healthcare providers across Europe to recruit diagnosed IMD patients, with a total of 24 participating centres. Recruitment is constrained by hospital-specific capacities and site-specific patient availability, making efficient cohort assembly non-trivial. In parallel, urine and plasma samples are collected and analysed at eight specialised metabolomics laboratories with differing platform capabilities, throughput, and costs. These analyses combine targeted and global metabolomics to characterise disease-specific metabolic profiles and support the reconstruction of IMD-specific genome-scale metabolic models.

The complexity and scale of this collaborative effort highlight the operational interdependencies that motivate the unified optimisation framework described in this study. By incorporating constraints on recruitment feasibility, laboratory capacity, biomarker coverage, cost, and clinical interest, the model ensures scalable, transparent, and clinically aligned study design across the network.

Selection of IMDs in the Recon4IMD project

The first step in the Recon4IMD recruitment strategy was to define a feasible list of inherited metabolic diseases for inclusion. Given the multidisciplinary scope of the project, selected IMDs needed to meet criteria spanning computational modelling, analytical feasibility, and clinical relevance. Each IMD had to have a known causal gene and a corresponding metabolic pathway represented in the latest genome-scale model of the human cell. The Virtual Metabolic Human (VMH) database was used to identify relevant genes and metabolites, excluding IMDs without clear genetic or metabolic biomarkers [17]. Metadata, including causal genes, biomarkers, and age-specific effect sizes, was obtained from the IEMbase database [13].

To ensure analytical feasibility, each participating metabolomics laboratory provided a list of metabolites measurable by their platforms. These were intersected with (i) IMD-associated biomarkers and (ii) metabolites present in the genome-scale model. This mapping was performed using MetaNetX, which integrates metabolic networks and biochemical pathway databases under the unified MNXref namespace [18].

Finally, recruitment feasibility was addressed using hospital-provided estimates of patient counts, ensuring that the selected IMDs had sufficient patient numbers to support statistically robust clinical studies.

Healthcare providers in Recon4IMD project

Healthcare providers (HCPs) serve as centres for recruiting patients with IMDs. Each HCP has a limited number of patients available per IMD and operates under a recruitment capacity determined by its facilities and human resources. In this case, most hospitals have a recruitment threshold of 40 patients, while three participating centres can recruit up to 250 patients. These constraints define the allowable values of the recruitment variable ziz_i, which represents the number of patients recruited per HCP.

A survey was conducted to collect clinical interest from each HCP. Each hospital submitted its preferred IMDs using IEMbase nosology code as the namespace or a whole ICIMD nosology group [19]. By mapping ICIMD nosology groups to their associated IEMbase IMDs, a clinical interest vector was constructed for each HCP at the IMD level. This clinical interest vector is incorporated into the integrated optimisation problem, as shown in Equation (9).

Effect size and cohort design by IMD

To date, no systematic study has established how many patients are needed to construct patient-specific genome-scale models across a large number of IMDs. From an optimisation perspective, defining quantified lower and upper bounds for cohort sizes is essential to balance statistical power, diagnostic certainty, and resource constraints. To address this gap, a scalable framework was developed to estimate appropriate cohort sizes using metabolomic data as a primary source of diagnostic information. This approach is intended as a rational starting point for guiding recruitment, particularly in the absence of disease-specific modelling precedents.

Diagnostic certainty is one of the key factors influencing the number of patients required for a valid clinical study. Studies of diseases that present clear metabolite biomarkers with a large effect size generally require fewer patients [5, 20, 21]. Conversely, conditions with less distinctive biomarkers and smaller effect sizes typically need larger cohorts to guard against misdiagnosis.

To formalise this relationship, we introduce the Diagnostic Confidence Score (DCS), a quantitative metric designed to estimate diagnostic reliability specifically from metabolomic data. The DCS integrates two components: the number of characteristic metabolite biomarkers and their respective effect sizes, as reported in IEMbase. This score reflects how distinguishable a disease is based on its metabolomic profile. Diseases with higher DCS values are assumed to be more confidently identifiable from metabolomic data alone and are therefore expected to require fewer patients for metabolic model construction. Conversely, lower DCS values indicate limited or ambiguous metabolomic signals and suggest a need for larger cohorts to mitigate diagnostic uncertainty.

The DCS is calculated using qualitative diagnostic annotations from the IEMbase database. Effect sizes are encoded in IEMbase using arrows, which are converted into numerical values to enable quantitative analysis: one arrow is assigned a value of one unit, two arrows a value of two units, and ambiguous ranges such as “normal to one arrow” are approximated as 0.5 units. The average of these values across all age-specific groups defines the quantified effect size for each IMD.

In parallel, we identify the number of feasible characteristic metabolite biomarkers, defined as those both reported in IEMbase and measurable by Recon4IMD partner laboratories. The Diagnostic Confidence Score is then defined as:

Diagnostic Confidence Score=[number of feasible characteristic metabolite biomarkers]×quantified effect size(15)\text{Diagnostic Confidence Score} = [\text{number of feasible characteristic metabolite biomarkers}] \times\text{quantified effect size} \tag*{(15)}

This score serves as a proxy for the diagnostic utility of metabolomic data in identifying a given IMD. To implement this framework in Recon4IMD, we stratified IMDs into three diagnostic confidence categories based on the DCS distribution: high (DCS >6> 6), moderate (3≤DCS≤63 \leq\text{DCS} \leq6), and low (DCS <3< 3). These thresholds were chosen based on the empirical distribution of DCS values calculated for all IMDs in IEMbase, as detailed in Supplementary Figure S1 and the accompanying explanation. The distribution is highly left-skewed, with a substantial proportion of IMDs exhibiting a DCS of zero, indicating no currently measurable characteristic biomarkers.

The cohort size for each IMD is case-specific, making it difficult to determine quantitatively the exact number of samples required for a clinical study. A recent study confirmed that sample sizes in rare disease studies vary widely, ranging from a few individuals to several hundred [22]. Several factors influence the determination of an appropriate sample size, including the expected effect size or difference between groups or time points, the degree of homogeneity among study participants, the acceptable risk of error, and the anticipated rate of participant attrition [5]. In addition to these, the overall scope of the project and the availability of resources such as funding, time, and personnel also play a significant role in shaping sample size decisions, as they can limit or expand the capacity to recruit and manage participants effectively. In this study, we projected an average of 10–15 patients for each disease cohort, which aligns with the Recon4IMD project plan to recruit approximately 1,900 diagnosed patients from a diverse group of IMDs. Initially, the defined DCS thresholds correspond to natural breaks in the distribution and were used to guide cohort size assignments: 5–10 patients for high-confidence IMDs, 10–15 patients for moderate-confidence IMDs, and 15–20 patients for low-confidence IMDs. Finally, the resulting cohort size recommendations were reviewed by clinicians involved in the project and adjusted on a case-by-case basis based on expert evaluation of each disease. This review step also reflects that the Diagnostic Confidence Score was used as a project-specific planning tool to guide cohort-size bounds, rather than as a validated diagnostic score or universal sample-size rule.

Adjusting patient estimates for practical recruitment limitations

Recruitment for clinical studies is often limited by operational and ethical challenges. Factors such as limited patient availability, infrequent clinic attendance, and inconsistent collection of diagnostic data can hinder the timely identification and enrolment of eligible participants. Additional barriers include those specific to paediatric studies (e.g. the need for dual parental consent in some countries), patient cancellations, reluctance to participate, and, in some cases, intentional avoidance of approach due to concerns about overburdening patients or jeopardising the clinician–patient relationship. These constraints reduce the realistically recruitable patient pool, even when a larger number may be theoretically eligible. To improve alignment with these limitations, we propose adjusting the model inputs to reflect not the total recorded patient counts at each site, but an estimated number of patients likely to participate. This modification increases the practical applicability of the optimisation framework by incorporating expected recruitment attrition into the planning assumptions, thereby enhancing both the feasibility and robustness of cohort allocation strategies.

Based on our experience, we anticipated that approximately 50% of estimated patients would ultimately provide a sample for the study. Therefore, we divided the estimated patient count matrix D\mathbf{D} by two to more accurately approximate the number of patients expected to be recruited. This adjustment can reduce the estimated recruitable pool below the nominal disease-specific cohort lower bound for some IMDs, making the integrated optimisation problem infeasible if the original cohort thresholds are imposed unchanged. We therefore applied a feasibility relaxation to the disease-specific cohort lower bounds, defining the minimum acceptable cohort size as a fixed fraction of the nominal target and rounding this value upward to the nearest integer. This ensured that the relaxation did not introduce fractional patient requirements. This approximation may vary across hospitals, depending on their registration procedures, the specific diseases involved, and the time span over which patient data are collected, as older records typically correspond to lower participation rates.

Metabolomic analysis in Recon4IMD project

Metabolomics is central to the Recon4IMD project, as it captures the biochemical consequences of metabolic dysfunction in IMDs. Both quantitative targeted and semi-quantitative global metabolomics are applied to each patient sample to ensure comprehensive biochemical characterisation, enabling integration with genome-scale metabolic models. Accordingly, the mathematical modelling of sample allocation, as described in the Methods section on sample allocation and logistics, applies to both global and targeted platforms within Recon4IMD.

Seven participating laboratories are co-located with clinical centres, forming hybrid sites capable of both patient recruitment and local sample analysis. This distributed configuration introduces logistical complexity into the sample allocation process, which must account for platform capabilities, analytical throughput, cost, and geographic proximity.

A detailed mathematical formulation of the integrated optimisation strategy is provided in Supplementary Methods. This includes platform capabilities, capacity constraints, cost modelling, and logistical penalties.

Metabolite identifier standardisation and interdisciplinary integration

The Recon4IMD project evaluated eight European metabolomic laboratories with varying analytical platforms, costs, and sample throughput capacities, highlighting the need for a unified metabolite identifier namespace to objectively compare platform capabilities and integrate metabolomics data into genome-scale models. Bridging experimental results with computational models requires harmonising identifiers used in analytical platforms with those in metabolic reconstructions. Multiple identifier systems exist for different use cases: the Human Metabolome Database (HMDB) offers human-centric annotations [23]; KEGG Compound IDs support pathway mapping [24]; PubChem CIDs provide extensive structural data [25]; SwissLipids offers lipid-specific, ontology-based annotations [26]; and ChEBI for ontology-driven annotation of chemical entities [27]. On the modelling side, the Virtual Metabolic Human (VMH) database serves as the reference for genome-scale reconstructions, linking its identifiers to databases like ChEBI, KEGG, and HMDB for interoperability [17]. Among these, ChEBI is notable for its structured ontology and its explicit representation of stereochemical and charged states, a critical feature given the limitations of many metabolomics platforms in resolving stereoisomers or isomeric lipids, especially in untargeted workflows [28, 29]. While stereoisomer identity can sometimes be inferred from biological context, such as the predominance of L-amino acids in plasma, multiple isomers often coexist, and ChEBI’s hierarchical ontology supports formal representation of this uncertainty [30]. Thus, ChEBI offers a robust, ontology-driven framework for metabolite identification in projects like Recon4IMD, enabling precise, semantically interoperable communication between experimental and computational teams [31].

Data privacy and anonymisation strategy

To comply with the General Data Protection Regulation (GDPR; Regulation (EU) 2016/679) and ethical obligations around patient confidentiality, we anonymised all sensitive data prior to analysis. Specifically, we withheld the estimated number of patients recruited at each healthcare provider and the allocation of samples to individual metabolomic laboratories. This step was taken to prevent indirect re-identification of patient subgroups or institutional profiles based on operational data.

To preserve analytical utility while ensuring confidentiality, we evaluated several anonymisation techniques, including k-anonymity [32], data aggregation, and synthetic data generation. We selected differential privacy as the most scientifically rigorous and formally grounded approach [33], offering provable privacy guarantees by injecting calibrated statistical noise. A privacy budget of ε=1.5\varepsilon= 1.5 was used, balancing privacy protection and model fidelity aligning with established practice for healthcare and public policy applications [34, 35].

Further technical details, including the mathematical formulation and implementation of the Laplace mechanism, are provided in the Supplementary Methods.

Model benchmarking

To assess the performance of the proposed cardinality-optimised patient recruitment and sample allocation framework, we conducted a structured benchmarking analysis against non-optimised baseline strategies. The evaluation was designed to separately examine recruitment performance and sample allocation performance while maintaining consistency in cohort definitions and input data. Full pseudocode for all baseline algorithms is provided in Supplementary Section ’Baseline Algorithms for Recruitment and Sample Allocation’.

Given the absence of an established integrated comparator for this combined planning problem, the proportional and greedy baselines were selected as realistic staged planning alternatives, representing both practical rule-based approaches and stronger optimisation-inspired greedy heuristics commonly used for marginal-gain and set-cover allocation problems. Recruitment performance was evaluated by comparing the cardinality-optimised model against two alternative recruitment strategies: a proportional heuristic recruitment baseline (Algorithm S1) [11] and a greedy recruitment baseline (Algorithm S2) [12]. The objective was to compare the total number of patients recruited under each method as the number of participating healthcare providers (HCPs) increased from single-centre inclusion to full multi-centre participation.

In the proportional heuristic recruitment strategy (Algorithm S1), patients were assigned to hospitals in proportion to site-specific patient availability. For each disease, weights proportional to the estimated patient counts at the selected hospitals were computed and used to distribute the disease-specific recruitment target, subject to hospital-level capacity constraints and the predefined lower and upper bounds on disease cohort sizes. If the initial proportional allocation did not satisfy the recruitment target for a disease, a rebalancing loop iteratively assigned remaining patients to hospitals with residual capacity and sufficient available patient counts. This strategy prioritises feasibility and local availability but does not incorporate global coordination across diseases or sparsity considerations.

In addition, a greedy recruitment baseline (Algorithm S2) was implemented. This approach iteratively assigns patients to disease–hospital pairs that maximise marginal recruitment gain while respecting hospital capacity constraints and patient availability limits. At each iteration, the feasible assignment yielding the largest immediate increase in recruited patients is selected. Although more structured than the proportional heuristic, this strategy remains locally greedy and does not optimise recruitment decisions jointly across diseases.

To evaluate sample allocation performance independently of recruitment differences, we first solved the integrated recruitment–allocation optimisation model under site-specific patient availability and laboratory capacity constraints. To establish a baseline comparison without the confounding effect of throughput limitations, the model was also solved under unconstrained laboratory capacity conditions. The resulting planned disease-specific patient totals, denoted by yy were then fixed and supplied to the allocation baseline algorithms to ensure that all allocation methods operated on an identical cohort.

In the allocation benchmarking phase, each required biomarker–disease measurement was treated as a demand for two analytical samples, corresponding to plasma and urine. Two allocation baselines were considered. The first, a heuristic disease-wise allocation strategy (Algorithm S3), applied a set-cover procedure independently for each disease to select analytical platforms that satisfied biomarker coverage requirements while limiting the number of platforms used. This approach treats diseases separately and does not exploit potential cross-disease efficiencies.

The second allocation baseline implemented a global greedy marginal set-cover strategy (Algorithm S4), following the classical greedy approximation framework for set covering. In this procedure, analytical platforms were selected sequentially based on maximising marginal biomarker coverage per unit cost across all diseases. This global selection phase was followed by a disease-specific allocation step restricted to the platforms chosen in the global phase. Although this method incorporates cross-disease information and cost-awareness, it remains greedy and does not guarantee global optimality.

Across all allocation experiments, identical biomarker requirements, cost parameters, and platform capability matrices were used. No re-optimisation of recruitment was performed in the baseline allocation cases, ensuring that performance differences reflected allocation strategy rather than cohort variation.

To assess robustness under operational constraints, progressively tighter sample throughput limits were imposed on both global and targeted platforms. Platform capacities were varied from 0 to 2000 samples per site in increments of 200, and feasibility was evaluated for the cardinality-optimised model as well as for the heuristic and greedy allocation baselines.

In addition to benchmarking performance and feasibility, we examined the model’s ability to manage trade-offs between competing objectives. Specifically, we analysed the interaction between alignment with hospital-specific clinical interests and minimisation of non-local sample shipment. A parametric sweep of the clinical interest weight λ\lambda was conducted over the range from 0 to 10,000. For each value of λ\lambda recruitment patterns and associated logistics penalties were recorded to characterise the resulting Pareto frontier.

Together, the HCP engagement scenarios, laboratory capacity sweep, and clinical-interest weight sweep provide sensitivity analyses for recruitment scale, operational capacity, and objective-weight calibration.

Detailed pseudocode for Algorithms S1–S4 is provided in Supplementary Section ’Baseline Algorithms for Recruitment and Sample Allocation’.

The hospital-provided estimates of patient counts used in this manuscript to assess the feasibility of clinical study recruitment and downstream research analysis contained no personal or patient-identifiable information and could not be linked to individuals. The data were fully anonymous and comprised only aggregate estimates of recorded cases for each disease or the numbers potentially available for recruitment at each hospital. Consequently, ethics approval and consent were not required for this recruitment-feasibility assessment, as it involved no human participants, no interventions, and no identifiable data.

The analyses reported here relate solely to recruitment-feasibility modelling and do not form part of the Recon4IMD Clinical Study. Within the Recon4IMD project, the terms governing collaboration and data transfer between all members, including collaborating hospitals, have been formally established through legal agreements. Under this framework, Recon4IMD members act as joint data controllers and processors.

The Recon4IMD Clinical Study is a separate prospective investigation that will involve patient recruitment, informed consent, biological sample collection, and the transfer of pseudonymised clinical data. Ethical approval for the clinical study is independently governed and has already been granted at several sites. At Tübingen approval was granted by the Ethics Committee at the Faculty of Medicine of the Eberhard Karls University and University Hospital Tübingen with Approval Number 371/2025B02. At Heidelberg approval was granted by the Ethics Committee of the Medical Faculty, Heidelberg University with Approval Number S-645/2024. In the United Kingdom approval was granted by the West Midlands Coventry and Warwickshire Research Ethics Committee with Approval Number 25/WM/0032. Recruitment at any site will only begin once the required local ethical approval has been obtained.

Data Availability

The clinical recruitment and metabolomic lab capability datasets analyzed during this study are available in the GitHub repository at https://github.com/Digital-Metabolic-Twin-Centre/clinical-study-optimisation. To protect participant privacy and institutional profiles, certain sensitive operational data, specifically the raw estimated patient counts per healthcare provider and the specific allocation of samples to individual laboratories, have been withheld in their raw form. These data were processed using differential privacy (ε=1.5\varepsilon= 1.5) to ensure anonymity while preserving analytical utility. Reference data for inherited metabolic diseases, causal genes, and metabolite biomarkers were sourced from public databases: the Virtual Metabolic Human (VMH) [17], IEMbase [13], and the International Classification of Inherited Metabolic Disorders (ICIMD) [19]. Any additional data supporting the findings of this study are available from the corresponding author upon reasonable request.

Code Availability

The complete implementation of the optimisation framework presented in this study is publicly available at https://github.com/Digital-Metabolic-Twin-Centre/clinical-study-optimisation. This repository includes a MATLAB Live Script (compatible with R2024a) providing end-to-end execution of the model, along with the anonymised datasets. The framework utilizes a cardinality optimisation function [8], available through the COBRA Toolbox [15] and requires the GUROBI solver [16]. Detailed documentation of mathematical formulations and testing routines is provided in the repository to ensure full reproducibility.

Acknowledgements

This study was funded by the European Union’s Horizon Europe Framework Programme (grant agreement No 101080997), the Swiss State Secretariat for Education, Research and Innovation (grant No 23.00232), United Kingdom Research and Innovation (grant Nos 10083717 and 10080153), and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy (CIBSS - EXC-2189 - Project ID 390939984) to L.H, and Czech grant projects (BBMRI_CZ LM2023033, RVO VFN64165).

Author Contributions

F.Z. conceived and designed the study, developed the mathematical framework, cleaned and analysed data, and drafted the manuscript. C.C. communicated with healthcare providers and compiled the data. T.D. helped with the IMD list and reviewed the manuscript. M.M. provided patient numbers, reviewed the IMD list, and provided clinical insight. J.T. helped with the IMD list and with the definition of the Diagnostic Confidence Score. J.M. analysed and compared the metabolomics data and communicated with the metabolomic labs. P.J., S.W.G., U.S., J.J.O’B., T.T., M.L.C., J.d.l.H., N.D., D.M., P.W., I.K., G.L.M., A.T.v.d.P., H.H.H., T.G.J.D., F.J.v.S., S.D.K.K., A.M.D., C.D.V., C.D.v.K., J.B.H., S.O’S., S.A.F., and M.S. contributed IMD data.

L.H., M.R.H.F., M.J.N., I.M.L.W.K.K., K.B.P.E., G.J.G.R., J.A.C., J.J.M.J., A.H., and T.H. provided metabolomic platform capabilities. S.R. contributed patient numbers and reviewed the manuscript. S.K. contributed IMD data, reviewed the manuscript, and provided patient participation statistics. I.T. provided the VMH data and contributed to the manuscript review. R.M.T.F. supervised the research, contributed to the development of the mathematical model, provided guidance on study design and methodology, contributed to the interpretation of results, and revised the manuscript.

Competing Interests

The authors declare no conflicts of interest.

References

  1. [1]Harden, M. & Friede, T. Sample size calculation in multi-centre clinical trials. BMC medical research methodology 18, 156 (2018).
  2. [2]Zurek, B. et al. Solve-RD: Systematic pan-European data sharing and collaborative analysis to solve rare diseases. European Journal of Human Genetics 29, 1325–1331 (2021).DOI
  3. [3]Hilgers, R.-D., FranzKönig & Senn, G. M. a. S. Design and analysis of clinical trials for small rare disease populations. Journal of Rare Diseases Research & Treatment 1 (2016).DOI
  4. [4]Lee, S. G., Chung, H.-J., Park, J. B., Park, H. & Lee, E. H. Harmonization of laboratory results by data adjustment in multicenter clinical trials. The Korean Journal of Internal Medicine 33, 1119–1128 (2018).DOI
  5. [5]Burmeister, E. & Aitken, L. M. Sample size: How many is enough? Australian Critical Care 25, 271–274 (2012).DOI
  6. [6]Tang, M. et al. Challenges of international oncology trial collaboration—a call to action. British Journal of Cancer 121, 515–521 (2019).DOI
  7. [7]Myers, B. A. et al. Recruitment effort and costs from a multi-center randomized controlled trial for treating depression in type 2 diabetes. Trials 20, 621 (2019).
  8. [8]Fleming, R. M. T. et al. Cardinality optimization in constraint-based modelling: Application to human metabolism. Bioinformatics 39, btad450 (2023).DOI
  9. [9]Kozłowska, E., Haltia, U.-M., Puszynski, K. & Färkkilä, A. Mathematical modeling framework enhances clinical trial design for maintenance treatment in oncology. Scientific Reports 14, 29721 (2024).DOI
  10. [10]Davila, D. N., Rodríguez, L. C. & Freire, S. L. Laboratory Samples Allocation Problem. In XoveTIC Congress 2018, 1189 (MDPI, 2018).
  11. [11]Chvatal, V. A Greedy Heuristic for the Set-Covering Problem. Mathematics of Operations Research 4, 233–235 (1979).DOI
  12. [12]Nemhauser, G. L., Wolsey, L. A. & Fisher, M. L. An analysis of approximations for maximizing submodular set functions—I. Mathematical Programming 14, 265–294 (1978).DOI
  13. [13]Ferreira, C. R., van Karnebeek, C. D. M., Vockley, J. & Blau, N. A proposed nosology of inborn errors of metabolism. Genetics in Medicine 21, 102–106 (2019).DOI
  14. [14]Deb, K. & Gupta, S. Understanding knee points in bicriteria problems and their implications as preferred solution principles. Engineering Optimization 43, 1175–1204 (2011).DOI
  15. [15]Heirendt, L. et al. Creation and analysis of biochemical constraint-based models using the COBRA Toolbox v.3.0. Nature Protocols 14, 639–702 (2019).DOI
  16. [16]Gurobi Optimization, LLC. Gurobi optimizer reference manual (2024).
  17. [17]Noronha, A. et al. The Virtual Metabolic Human database: Integrating human and gut microbiome metabolism with nutrition and disease. Nucleic Acids Research 47, D614–D624 (2019).DOI
  18. [18]Moretti, S., Tran, V. D. T., Mehl, F., Ibberson, M. & Pagni, M. MetaNetX/MNXref: Unified namespace for metabolites and biochemical reactions in the context of metabolic models. Nucleic Acids Research 49, D570–D574 (2021).
  19. [19]Ferreira, C. R., Rahman, S., Keller, M., Zschocke, J. & Group, I. A. An international classification of inherited metabolic disorders (ICIMD). Journal of Inherited Metabolic Disease 44, 164–177 (2021).
  20. [20]Dobbin, K. K. & McShane, L. M. Sample size methods for evaluation of predictive biomarkers. Statistics in Medicine 41, 3199–3210 (2022).DOI
  21. [21]Serdar, C. C., Cihan, M., Yucel, D. & Serdar, M. A. Sample size, power and effect size revisited: Simplified and practical approaches in pre-clinical, clinical and laboratory studies. Biochemia Medica 31, 27–53 (2021).
  22. [22]Hee, S. W. et al. Does the low prevalence affect the sample size of interventional clinical trials of rare diseases? An analysis of data from the aggregate analysis of clinicaltrials.gov. Orphanet Journal of Rare Diseases 12, 44 (2017).DOI
  23. [23]Wishart, D. S. et al. HMDB 5.0: The Human Metabolome Database for 2022. Nucleic Acids Research 50, D622–D631 (2021).DOI
  24. [24]Kanehisa, M., Furumichi, M., Sato, Y., Kawashima, M. & Ishiguro-Watanabe, M. KEGG for taxonomy-based analysis of pathways and genomes. Nucleic Acids Research 51, D587–D592 (2023).DOI
  25. [25]Kim, S. et al. PubChem 2023 update. Nucleic Acids Research 51, D1373–D1380 (2023).DOI
  26. [26]Aimo, L. et al. The SwissLipids knowledgebase for lipid biology. Bioinformatics 31, 2860–2866 (2015).DOI
  27. [27]Hastings, J. et al. ChEBI in 2016: Improved services and an expanding collection of metabolites. Nucleic Acids Research 44, D1214–D1219 (2016).DOI
  28. [28]Oliveira, R. V., Simionato, A. V. C. & Cass, Q. B. Enantioselectivity Effects in Clinical Metabolomics and Lipidomics. Molecules 26, 5231 (2021).DOI
  29. [29]Addepalli, R. V. & Mullangi, R. A concise review on lipidomics analysis in biological samples. ADMET & DMPK 9, 1–22 (2020).DOI
  30. [30]Novoa-del-Toro, E. M. & Witting, M. Navigating common pitfalls in metabolite identification and metabolomics bioinformatics. Metabolomics 20, 103 (2024).
  31. [31]Poupin, N. et al. Improving lipid mapping in Genome Scale Metabolic Networks using ontologies. Metabolomics 16, 44 (2020).DOI
  32. [32]Sweeney, L. K-ANONYMITY: A MODEL FOR PROTECTING PRIVACY. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 10, 557–570 (2002).DOI
  33. [33]Dwork, C., McSherry, F., Nissim, K. & Smith, A. Calibrating Noise to Sensitivity in Private Data Analysis. In Halevi, S. & Rabin, T. (eds.) Theory of Cryptography, 265–284 (Springer, Berlin, Heidelberg, 2006).
  34. [34]Dwork, C. & Roth, A. The Algorithmic Foundations of Differential Privacy. Foundations and Trends® in Theoretical Computer Science 9, 211–407 (2014).
  35. [35]Erlingsson, Ú., Pihur, V. & Korolova, A. RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response. In Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security, CCS ’14, 1054–1067 (Association for Computing Machinery, New York, NY, USA, 2014).

Paper details

Contents