Genetics by genetic algorithm: Defining an ideal platelet donor population to support patients with HLA‐mediated alloimmune refractoriness

Abstract

Background

Immune factors underlie approximately 20% of patients exhibiting platelet transfusion refractoriness (PTR). PTR can be mitigated by supplying HLA‐A and B compatible platelets to these patients. Since 2019, Canadian Blood Services has managed a program to provide compatible platelets to alloimmunized patients. This study evaluates donor pool size and ethnic diversity options to serve Canada's population.

Study Design and Methods

We adapted simulation‐based stem cell matching methods by generating simulated patients and matching them to existing and simulated donors using a Python algorithm. Recruitment scenarios included census‐aligned, single‐group, and meta‐heuristic–optimized distributions to enhance match rates and equity.

Results

More typed donors increase match rates, but rates vary significantly by ethnicity. Non‐Black patients achieved 95% coverage with 13,000–18,000 new donors, whereas Black patients required ~55,000 new donors. Single‐ethnicity recruitment improved the targeted group's rate but reduced overall match rates and disadvantaged other populations. A Pareto‐optimal donor mix—augmenting Black, Hispanic, Asian Pacific Islander, and Native American/First Nations donors while modestly reducing White typing—preserved or improved rates but yielded only modest gains for Black patients. A balanced strategy further prioritizing Black and Hispanic donors with slight reductions in White, Asian, and Native American/First Nations typing achieved improvements for the Black population with minimal losses for others.

Discussion

While Pareto‐optimal allocation enhances overall efficiency, targeted trade‐offs, focusing on underrepresented groups, are essential to correct persistent disparities and achieve equitable HLA‐matched platelet availability.

Abbreviations

API

Asian and Pacific Islander

CBS

Canadian Blood Services

DEAP

Distributed Evolutionary Algorithms in Python

GA

Genetic Algorithm

HIS

Hispanic

HLA

Human Leukocyte Antigens

HPA

Human Platelet Antigens

NAM

Native American/First Nation

NMDP

National Marrow Donor Program

OTH

Other

PTR

Platelet Transfusion Refractoriness

1. INTRODUCTION

Platelet transfusion refractoriness (PTR) is defined as a poor increment in the platelet count post‐transfusion. ¹ Patients exhibiting platelet refractoriness remain at risk of severe bleeding due to underlying thrombocytopenia. PTR is attributed to immune factors in about 20% of cases. ² Platelets express Human Leukocyte Antigens (HLA) Class I (HLA‐A and HLA‐B), Human Platelet Antigens (HPA), and ABO antigens, among others. Alloimmune refractoriness occurs most commonly due to antibodies against Class I Human Leukocyte Antigens (HLA). These antibodies lead to rapid clearance of transfused platelets. However, alloimmune PTR can often be circumvented by providing platelets that are either matched at the HLA A and B loci or that lack the specific antigen specificities of existing recipient antibodies. ³

Canadian Blood Services (CBS) maintains a national program to provide HLA and HPA selected platelets for alloimmunized patients. The medical complexity and ethnic diversity of the patient population supported by this program are increasing, motivating a review of the ideal size and ethnic composition of the apheresis platelet donor population to sustain the HLA/HPA platelet program. This study evaluates donor pool size and ethnic diversity required to serve Canada's population (outside of the Province of Quebec), focusing specifically on HLA Class I typing and matching.

Previous studies have demonstrated that the best results are associated with perfect matches. Further, there are reports of models designed to determine the minimum donor pool size required to support a population of platelet‐refractory patients. Bolgiano et al. ⁴ provide a seminal work in which donor HLA frequencies are determined from a sample of cadaveric kidney donors and matched against patient profiles. They estimated that between 1000 and 3000 fully typed platelet donors were necessary to meet most demands with one or fewer mismatches.

Bub et al. adopt a data search approach to calculate donor‐base size for matched platelets. A sample of 150 patient records is matched against a potential donor population of 65,000 typed individuals, with both populations being drawn from within Brazil. Counting only perfect matches, they estimate that between 6000 and 32,000 typed donors are necessary to find between 1 and 5 donors for 80% of the patient population. ⁵

While the number of reports in the literature regarding donor pool sizing for HLA‐matched platelets is small, the problem can be considered a more restricted version of the hematopoietic stem cell matching problem. Extensive literature on stem cell matching and donor pool sizing exists. ⁶ , ⁷ , ⁸ , ⁹

Our research group developed an integer programming model to select recruits according to ethnicity to meet stem cell matching requirements in Canada. ¹⁰ A later study employed the concept of simulated patients and donors to determine the impact of increased recruiting efforts within ethnic groups on overall patient match rates. ¹¹

The goal of the present study is to determine the donor pool size and optimal recruiting efforts required, by ethnicity, to support the HLA‐selected platelet program in Canada and to promote equity among patient groups of varied ethnicities.

2. STUDY DESIGN AND METHODS

Approximately 9000 individuals are on the platelet donor registry maintained by Canadian Blood Services, with approximately 3700 persons who have donated in the past 2 years and are considered active. Apheresis platelet collection accounts for approximately 30% of the annual platelet inventory at CBS. All donors enrolled in the apheresis platelet collection program are genotyped for HLA Class I A and B via sequence‐specific oligonucleotide testing. Donors have the option to self‐report their ethnicity, based on a pre‐defined list of options at the time of donation. The self‐reported ethnicity of typed donors in the registry appears in Table 1, aggregated by broad ethnic category to fit available data sources, where beyond Black and White API = Asian and Pacific Islander, HIS = Hispanic, NAM = Native American/First Nation, and OTH = Other.

TABLE 1.

Self‐reported ethnicities for typed donors aggregated by broad ethnic category census averages.

Ethnicity	Black	API	White	HIS	NAM	OTH
Self‐Reported	0.42%	6.38%	86.8%	0.05%	2.1%	3.55%
Canadian Census	4.3%	19.3%	68.7%	2.5%	6.1%	–

As shown listed in Table 1, the composition of self‐reported ethnicity of platelet donors is discordant from the latest Canadian census data, with over‐representation of White donors and under‐representation of all other ethnic categories. ¹² This mirrors the demographics of Canada's whole blood donor population. However, as the Canadian population becomes more diverse, it will be necessary to modify methods of platelet donor recruitment to better account for ethnic diversity and to increase the chances all ethnic groups have equitable representation within the donor base.

2.1. Simulated donors

To estimate the impact of donor recruitment strategies, differing numbers of donors, with varying ethnic distributions, are recruited in the simulation. Donor recruitment in this study involves creating a specified number of donors (0 to 10,000), each of whom is assumed to be typed at the HLA‐A and ‐B loci and is added to the existing registry of 3700 active persons.

Each newly created donor is assigned an ethnicity according to a specified ethnicity distribution. Then, using ~A ~ B haplotype frequency data from US sources, ¹³ a pair of haplotypes is assigned to the donor. This creates a simulated donor with a complete HLA Class I phenotype.

2.2. Simulated patients

In a fashion similar to that of donors, 10,000 simulated patients are created by selecting an ethnicity from an empirical distribution based on the Canadian census and assigned a pair of ~A ~ B haplotypes, using a distribution of haplotype frequency specific to the patient's assigned ethnicity.

2.3. Simulated matches

Patients are then matched to donors (both newly created and existing) using an algorithm coded in Python. The algorithm is a re‐coding of methods reported in. ¹¹ In the results reported below, only perfect HLA matches (4 out of 4) are provided. However, lower quality matches can also be reported. The Python code and a skeleton MS‐SQL Server database to drive the model are available online.

2.4. Experiments

Experiments were run with increasing numbers of newly typed donors, ranging from 0 to 10,000. The ethnic distribution of newly typed donors was first matched to that of the Canadian census (as shown in Table 1) and results recorded. Next, the ethnic distribution of new donors was set to be entirely one of the five ethnic categories used by the US National Marrow Donor Program [NMDP] (AFA, API, CAU, HIS, or NAM), and the experiments were repeated. ¹³ Then, using a meta‐heuristic, a set of experiments was run to identify a near‐optimal ethnic distribution for a cohort of 10,000 newly typed donors. Finally, using the same meta‐heuristic, a recruiting cohort was identified that promoted match equity among ethnic groups.

3. RESULTS

In the first set of experiments, an increasing number of newly typed donors is added to the existing matched platelet donor registry under the assumption that the ethnic distribution of donors follows that of the Canadian census (see Table 1). Additional newly typed donors ranged from 0 to 10,000 in the experiments. In each experiment, 10,000 patients are generated and matched against the total registry of donors, which consists of 3700 currently active, typed donors along with the newly typed donors. Results appear in Figure 1.

FIGURE 1.

Graph of match rates, by ethnic origin of the patient, versus additionally typed donors, assuming both newly typed donors and patients are created in proportion to the ethnic distribution of the Canadian population.

Results show that match rates increase, among all ethnic groups, as the number of newly typed donors increases. An analysis of variance shows that donor match rate is significantly affected by both the number of newly typed donors and the ethnicity of the recipient. Using Tukey's test, it can be shown that the match rate of the Black population differs significantly from all other ethnic groups across all levels of additional donor typing. Empirically, differences in match rates may be addressed either by typing many new donors or by expending effort to enrich typing among Black donors.

3.1. Addressing match rate inequality via additional donor typing

Using the data from Figure 1, non‐linear regression models of the form $R_e=c{D}^{x_e}$ were developed, where R _e is the match rate for ethnic group e, D is the total number of newly typed donors, and x _e is a power rate value specific to the group. Using the power equations derived from the data (see Table 2), the number of newly typed donors needed to obtain a 95% match rate for each ethnic group can be estimated.

TABLE 2.

The constant (c) and power (x _e ) to fit a function of form R _e = ${\mathit{cD}}^{x_{\mathrm{e}}}$ to the data in Figure 1.

	c	x e	Match rate given 10,000 newly typed	Newly typed required for 95% match rate
Black	0.016	0.372	49.43%	55,257
API	0.055	0.301	85.96%	12,929
White	0.180	0.175	88.46%	13,696
HIS	0.060	0.283	78.11%	17,249
NAM	0.161	0.184	85.64%	15,551

Note: Also shown is the simulated match rate at 10,000 newly typed donors and the number of typed donors necessary to achieve a match rate >95% for each ethnic.

3.2. Addressing match rate inequality via targeted donor typing

Non‐Black groups can be assured of a match 95% of the time, if the number of newly typed donors is between 13,000 and 18,000. However, to achieve a similar match rate for Black patients, a much larger typing effort of approximately 55,000 individuals is required. An alternative strategy might be to differentially type donors from some ethnic groups to increase match rates. At the extreme, a strategy for increasing the match rate for patients from the Black population might be to type only Black donors.

In Figure 2 match rates versus extra typing are shown, under the assumption that all newly typed donors are drawn from the Black population. (Note that patients continue to be drawn from an ethnic distribution reflective of the current Canadian census.) As may be seen, the results are not Pareto optimal—while Black match rates increase (+31.2% at 10,000 newly typed), the rate for all other groups (−27.3% at 10,000 newly typed) and for Asian and Pacific Island patients specifically (−44.8% at 10,000 newly typed) decline.

FIGURE 2.

Graph of match rates, by ethnic origin of the patient, versus additionally typed donors, assuming that newly typed donors are all drawn from the AFA population.

Similar changes in match rates were noted under the assumption that a single ethnic group is typed, and 10,000 donors are added to the registry. Results appear in Table 3.

TABLE 3.

Changes in match rates for ethnic groups assuming newly typed donors are drawn from a single ethnic group and that 10,000 newly typed donors are added to the registry.

Community	New donor recruitment strategy
	Black only	API only	White only	HIS only	NAM only
Black	31.16%	−20.66%	−12.85%	−9.96%	−10.75%
API	−44.76%	4.01%	−22.37%	−21.06%	−20.42%
White	−20.95%	−21.86%	−1.32%	−5.79%	−5.99%
HIS	−20.61%	−16.96%	−5.04%	6.25%	−8.87%
NAM	−22.68%	−17.96%	−2.30%	−0.94%	−1.88%
Non‐Black	−27.25%	−13.20%	−7.76%	−5.39%	−9.29%
Total	−15.57%	−14.69%	−8.78%	−6.30%	−9.58%

Single ethnic group recruiting generally increases the match rate for patients originating from that same ethnic group (values along the diagonal). Note that recruiting only White or Native American/First Nations donors yields a small negative change, not statistically different from 0, in match rates for patients from these groups. Typing donors from only one ethnic group, however, always yields a net decrease in overall match rates and, in each case, results in a decrease in match rate for at least one different ethnic group. Thus, a more sophisticated method for finding a typing strategy is needed.

3.3. Searching for a typing strategy

If the number of newly typed donors is fixed (at 10,000 in the experiments that follow), it is possible to use the simulated donor creation and matching algorithm to search for an ethnic distribution that best promotes equity while minimizing decreases to any ethnic group.

Thus, an objective function was defined that evaluates the fitness of a particular donor typing distribution by comparing match rates for that distribution against the match rates expected if donors were typed according to the distribution of ethnicity within the Canadian population.

$$\mathit{Max}:K+\sum _{e\in E}\left[\mathit{Max}\left(0,R_e-R_{0e}\right)-10*\mathit{Max}\left(0,R_{0e}-R_e\right)\right],$$ (1)

where R _e is the match rate for ethnic group e where $e\in \left\{\text{Black}\mathit{API},\mathit{\text{White}},\mathit{HIS},\mathit{NAM}\right\}$. R _0e is the match rate for ethnic group e if donors are type according to the distribution of ethnicity within the Canadian population. K is an arbitrary constant to ensure a positive valued function (as required for the meta‐heuristic method employed).

The objective function of this search evaluates the difference between the observed match rate (R _e ) and the default match rate (R _0e ) for each ethnic group. Increases in match rates (R _e − R _0e ) are weighted by 1, while decreases in match rates (R _e − R _0e ) are weighted by 10 to discourage decreases in match rates. Since the absolute values are taken in both terms of the objective function, only non‐negative values are considered and thus one or the other term, but not both, can be non‐zero for any ethnic group.

The search for a good solution involves repeated calls to the matching algorithm using different values for the assumed ethnic distribution of the donors being newly typed and added to the donor registry. A trial solution thus comprises five floating‐point values—each representing the fraction of the donor typing assigned to a particular ethnic group:

$$x_{\mathit{\text{trial}}}=\left\{p_{\mathit{\text{Black}}},p_{\mathit{API}},p_{\mathit{\text{White}}},p_{\mathit{HIS}},p_{\mathit{NAM}}\right\},$$

where p _e is the fraction of the donor typing assigned to ethnic group e. Note that, by definition, ${\sum }_{e\in E}{p}_e=1$.

Several types of search strategies could be employed to find a good typing distribution, but because the function is not analytically defined, calculus‐based methods cannot be employed. Instead, a meta‐heuristic solution method is adopted. Since each trial solution appears in the form of a vector of probabilities, a genetic algorithm was adopted to search for good solutions.

According to Katoch, Singh‐Chauha, and Kumar ¹⁴ a genetic algorithm (GA) is a heuristic search algorithm derived from the analogy of a biological evolution process, employing “survival of the fittest.” The elements of GA are chromosome representation, fitness selection, and change operators, each based on genetic analogues. A chromosome, or trial solution, is comprised of several elements (“loci”), each of which has a value (p _e in our problem). Any chromosome can be evaluated according to the fitness function (Equation 1) using the simulated match program to determine the rates.

The algorithm starts with the generation of an initial population of trial solutions of a fixed size. In the first generation of the population, the values stored in the loci of the trial solutions are generated at random. Each trial solution within the population is then evaluated and ranked. This is the end of the initial generation. The population of solutions is then updated, either through crossover or mutation operations. Crossover is an operation in which pairs of promising solutions— typically those with higher fitness scores— are combined. This involves selecting a crossover locus within the chromosome of the trial solutions, which splits each solution into two parts. A new pair of trial solutions (children) is then created by combining one component from each parent. Chromosomes may also be subject to a mutation process, where some number of alleles are modified within the chromosome. Mutation sites are selected at random, and updated values are created according to a mutation rule. Once all trial solutions within a population at a particular generation are evaluated, the solutions are ranked, and lower quality solutions are eliminated from the population. The process then repeats. Thus, through successive generations, solution values evolve and generally improve. After a specified number of generations have evolved, the algorithm stops, and the best solution is returned. ¹⁵

In this search, a Python library, DEAP (Distributed Evolutionary Algorithms in Python), was used as the basis for genetic algorithm implementation. ¹⁶ DEAP implements the GA algorithm in a multi‐threading environment, which allows multiple chromosomes to be evaluated simultaneously. Additionally, DEAP provides simple mechanisms to implement custom mutation and crossover operators. A crossover operator was developed to provide for a single point swap of allele values in which the sum of probabilities within the newly constructed chromosomes was normalized to 1. Similarly, a mutation operator was constructed that also ensured the sum of probabilities within the chromosome sums to 1 after randomly mutating one locus. And, of course, when creating an initial population, a normalization routine was run to enforce the constraint that the sum of probabilities must equal 1.

The parameters for the GA were tuned by altering the objective function in Equation (1), such that positive and negative deviations from the default match rates are penalized equally. A constant of 100 is added to the equation to ensure that objective function values are positive and that the function can be maximized, as required for the eaSimple algorithm in DEAP.py.

$$\mathit{Max}:100-\sum _{e\in E}\left|\left(R_{0e}-R_e\right)\right|.$$ (2)

Parameters were then varied in a set of experiments in which the number of generations was altered between 10 and 20, and the initial population size was varied between 10 and 20. Good results can be obtained when a population size of 15 evolving over 15 generations is run.

Search results returned by the GA using the tuned parameters appear in Table 4 along with a repeat of the match rates identified by using single ethnic group typing.

TABLE 4.

Changes in match rates for ethnic groups assuming newly typed donors from Table 4 along with match rates for the best solution found by the genetic algorithm.

Community	New donor recruitment strategy
	Black only	API only	White only	HIS only	NAM only	GA solution
Black	31.16%	−20.66%	−12.85%	−9.96%	−10.75%	~0
API	−44.76%	4.01%	−22.37%	−21.06%	−20.42%	4.03%
White	−20.95%	−21.86%	−1.32%	−5.79%	−5.99%	~0
HIS	−20.61%	−16.96%	−5.04%	6.25%	−8.87%	~0
NAM	−22.68%	−17.96%	−2.30%	−0.94%	−1.88%	5.36%

Note: Note that 10,000 newly typed donors are assumed in all runs.

Results from Table 4 show a Pareto optimal solution can be found—one in which match rates increase for at least one ethnic group and do not decrease for any other group—by differentially selecting donors for typing based on ethnic origin.

The solution requires that Black, White, and Native American/First Nations donors be recruited at less than their distribution in the population, while over‐recruitment is necessary for Asian and Pacific Islander and Hispanic donors. See Table 5. Note that match rates for the Black population, while not significantly different from typing at the census distribution, remain statistically lower than those of other ethnic groups. However, this inequity can be addressed if we relax the constraint on Pareto optimality and minimize the match rate differences between ethnic groups.

TABLE 5.

Genetic algorithm solution showing the suggested donor typing rate and the resulting match rate expected, should such a policy be implemented.

	Black	API	White	HIS	NAM
Census distribution	4.3%	19.3%	68.7%	2.5%	6.1%
Current typing	0.42%	6.38%	86.8%	0.05%	2.1%
Suggested typing	1.0%	35.7%	40.3%	18.5%	4.5%
Match rate	46.00%	90.00%	88.00%	79.00%	91.00%

Note: Note that 10,000 additionally typed donors are assumed.

Table 6 shows a solution that minimizes the variance between the overall match rate (80%) and the match rate of the individual groups. The results suggest even more emphasis on two ethnic groups (Black and Hispanic) that are under‐represented in the matched donor registry and decreased emphasis on the White community, which is currently over‐represented relative to the census. Decreases in typing for Asian and Pacific Islanders and Native American/First Nations communities are also suggested, indicating that a more even distribution of match rates can be achieved by emphasizing recruitment elsewhere.

TABLE 6.

Solution showing suggested distribution for typing new donors to minimize variance for match rates across ethnic groups.

	Black	API	White	HIS	NAM
Census distribution	4.3%	19.3%	68.7%	2.5%	6.1%
Current typing	0.42%	6.38%	86.8%	0.05%	2.1%
Suggested typing	44.33%	3.83%	33.99%	16.03%	1.82%
Match rate	75.00%	74.00%	85.00%	81.00%	85.00%

Note: Note that 10,000 additionally typed donors are assumed.

4. DISCUSSION

In this study, the simulation‐based methods of Blake et al. ¹¹ were adapted to determine the donor pool size and ethnicity‐based recruitment strategy needed to optimally support the HLA‐selected platelet program in Canada and to promote equity among patient ethnic groups. Currently, the program utilizes both matched and antigen‐negative strategies when searching for suitable donors. Using the simulated patient and donor match algorithm, we found that increasing the number of typed donors results in an across‐the‐board increase in match rates for all ethnic groups in Canada, reducing the need for providing antigen‐negative donors. However, the number of donors required to achieve 95% coverage was significantly larger for the Black community than for other groups.

Accordingly, we conducted a set of experiments to search for an appropriate target for the ethnic distribution of newly typed donors while holding the typed donor registry at 10,000 additional persons. We employed a genetic algorithm to suggest distributions that are tested for fitness via the simulated patient–donor match program. This algorithm suggests that increasing donors from Asian and Pacific Islander and Hispanic communities, accompanied by a small decrease in White and Native American/First Nations donor recruiting, would provide the best overall match rate for the Canadian patient population. This solution is Pareto optimal in that the match rates for all ethnic groups either remain constant or increase. Nevertheless, the suggested typing profile retains match rates for the Black population that are statistically lower than those of other ethnic groups. To promote more equitable match rates across all ethnic groups, we identified an equity‐based typing approach that differentially selects currently under‐represented groups in the registry. This rate would increase match rates among Black patients with only small decreases in the match rates for Asian and White populations but would result in a more uniform match rate across all patient populations.

We note that simulated patients and donors in this study are all assumed to be members of one of the six broad ethnic groups defined in the US NMDP data set. Individuals of mixed ethnicity are not included, due to the lack of haplotype frequency data for such groups. Accordingly, overall match rates for the Canadian population will likely be lower than specified by the model. We note, therefore, the need for HLA haplotype frequency data tailored for populations with mixed ethnicity.

This study introduces a method for evaluating donor‐recipient match strategies using simulated patient and donor profiles. The framework supports flexible testing of recruitment and allocation policies under controlled conditions. A search capability based on meta‐heuristics was implemented to explore strategies that optimize different objective functions.

Using this approach, we identified two recruiting strategies for the Canadian population. One strategy balances and identifies a Pareto optimal solution across the entire population; the other reduces variation in match rates across ethnic communities. These strategies reflect different priorities and illustrate how the method can be adapted to support policy decisions.

Our method is general and may be applicable in other jurisdictions and with different products.

CONFLICT OF INTEREST STATEMENT

The authors have disclosed no conflicts of interest.

Blake JT, Rickards N, Mack J, Seftel M, Gupta A. Genetics by genetic algorithm: Defining an ideal platelet donor population to support patients with HLA‐mediated alloimmune refractoriness. Transfusion. 2025;65(12):2398–2405. 10.1111/trf.18446

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available in NMDP Registry Haplotype Frequencies at https://frequency.nmdp.org/. These data were derived from the following resources available in the public domain: HLA‐A~B, https://frequency.nmdp.org/jsp/haploTypeFrequencies.jsp.