Optimizing Microsatellite Marker Panels for Genetic Diversity and Population Genetic Studies: An Ant Colony Algorithm Approach with Polymorphic Information Content

Simple Summary

Microsatellite markers are widely used molecular markers for genetic studies, but choosing the right set involves a challenging trade-off between effectiveness and cost. The research aims to enhance the widely used ant colony optimization algorithm by integrating marker effectiveness indicators. By considering the genetic properties of the markers such as the polymorphic information content, the study seeks to determine the suitable way to select a reduced set of microsatellites. The approach addresses the accuracy–cost trade-off, aiding genetic assessments, breeding, and conservation efforts with cost-effective solutions. This research provides valuable insights into real-world genetic studies, including breeding programs and conservation initiatives.

Abstract

Microsatellites are polymorphic and cost-effective. Optimizing reduced microsatellite panels using heuristic algorithms eases budget constraints in genetic diversity and population genetic assessments. Microsatellite marker efficiency is strongly associated with its polymorphism and is quantified as the polymorphic information content (PIC). Nevertheless, marker selection cannot rely solely on PIC. In this study, the ant colony optimization (ACO) algorithm, a widely recognized optimization method, was adopted to create an enhanced selection scheme for refining microsatellite marker panels, called the PIC–ACO selection scheme. The algorithm was fine-tuned and validated using extensive datasets of chicken (Gallus gallus) and Chinese gorals (Naemorhedus griseus) from our previous studies. In contrast to basic optimization algorithms that stochastically initialize potential outputs, our selection algorithm utilizes the PIC values of markers to prime the ACO process. This increases the global solution discovery speed while reducing the likelihood of becoming trapped in local solutions. This process facilitated the acquisition of a cost-efficient and optimized microsatellite marker panel for studying genetic diversity and population genetic datasets. The established microsatellite efficiency metrics such as PIC, allele richness, and heterozygosity were correlated with the actual effectiveness of the microsatellite marker panel. This approach could substantially reduce budgetary barriers to population genetic assessments, breeding, and conservation programs.

1. Introduction

Microsatellite repeats, also known as simple-sequence repeats, are abundant and highly polymorphic in numerous eukaryotic genomes. They represent a class of DNA markers with repeat sequences ranging usually from mononucleotides to hexanucleotide repeats. Perfect repetitions, interrupted repeats, or combinations with other repeat types are possible occurrences. Biparentally inherited nuclear DNA microsatellites enable diverse applications, including population characterization, origin determination, hybrid identification, and the assessment of inbreeding levels. Consequently, while genome-wide single-nucleotide polymorphisms (SNPs) are frequently employed in genetic studies related to populations, forensics, conservation, and evolution, it is worth noting that microsatellite genotyping may offer a greater degree of informativeness compared to biallelic SNP genotyping in several species. This heightened informativeness arises from the fact that microsatellites represent mutational hotspots, characterized by elevated levels of polymorphism and a larger allelic diversity within diverse populations [1,2,3,4]. The high polymorphism and Mendelian inheritance of microsatellites make them a good choice, with significant impacts on breeding programs and conservation efforts. The global utilization of microsatellite markers in local laboratories with low-cost investment is a practical alternative to SNP genotyping, which requires advanced equipment and technology. However, the number of suitable microsatellite loci, which ranges from 10 to 30, may vary depending on the study field and research group. To measure the level of genetic variation and inbreeding in indigenous chickens, 15–30 loci derived from FAO reference markers were used [5]. An interpretation bias arises when comparing data on diversity and identification owing to the utilization of a large, non-optimized marker panel. However, the use of such a panel does not guarantee accurate results and can lead to a significant waste of human and financial resources, ultimately resulting in biased outcomes. The precision and accuracy of every downstream process following genotyping are mainly dependent on the effectiveness of the microsatellite panel. Admittedly, while a larger number of loci logically provides more genetic information on a population, researchers must consider a compromise between result accuracy and cost-effectiveness by accounting for the margin of error and defined accuracy criteria.

The widely used ant colony optimization (ACO) algorithm is a heuristic, population-based, and bioinspired optimization method for solving combinatorial problems [6]. This concept was proposed by Colorni et al. [7]. By leveraging the inherent behaviors observed in ant colonies, the ACO algorithm aims to determine the optimal solution by considering a set of constraints or costs [8]. The selection of an optimal microsatellite panel is driven by the intricate relationship between the utilized loci and the inferred result, leading to the categorization of the problem as nonlinear programming [9]. Solving these problems becomes computationally aspirational, even when dealing with a reasonable number of microsatellite markers, owing to the existence of multiple discrete decision variables [10]. Similar methods have been proposed to address these problems, including the genetic algorithm [11], particle swarm optimization [12], traveling salesman [13], and ant colony algorithm [8], which correspond to the ACO algorithm. In each method, the resource consumption and underlying logic differ; however, they all display remarkable flexibility in resolving optimization problems across various research domains [14]. These algorithms identified suitable microsatellite marker sets without relying on prior genetic knowledge. However, owing to the stochastic nature of metaheuristic algorithms, a local solution, characterized by high accuracy, but not necessarily the optimal accuracy among all possibilities, may be discovered, which could be distant from the global solution [15].

In this study, we aimed to elucidate the critical accuracy/cost trade-off dilemma in population genetics research projects. Here, rather than using a raw heuristic optimization algorithm, the effect of incorporating polymorphic information on the algorithm’s performance was explored. We hypothesized that integrating a relevant effectiveness indicator of a marker set into the ACO algorithm can lead to valuable findings such as reduced computational time and improved accuracy in identifying the optimal solution. When selecting the optimal microsatellite panel, the accuracy indicator was used as the cost function to be maximized [16]. Several approaches have considered polymorphic information content (PIC) [17], matching probability [18], and gene variability [19] as accuracy indicators for microsatellite panels. Additionally, a genetic distance matrix was used to provide useful information for population structure estimation using a reduced set of microsatellites [20]. By conducting a comparative analysis, the impact of incorporating PIC as a decision variable in the algorithm was evaluated. Our approach can help address budgetary barriers to population genetic assessments, breeding, and conservation programs.

2. Materials and Methods

2.1. Refining an Intriguing Algorithm for Microsatellite Marker Selection

The microsatellite marker selection problem is characterized as a combinatorial search problem, where there is a search space S and a cost function f that must be minimized [10]. The search space S comprises all possible subsets of markers, totaling 2^k potential solutions for k loci. Each subset was represented by a binary vector I = [i₁, i₂, …, i_n], where i ∈ {0;1} indicated whether a specific microsatellite was included in the marker panel or not. The accuracy of a microsatellite marker panel on a given genotype dataset was quantified using the cost function f. The cost function f was determined by comparing the average genetic distance (AGD) between the full set of markers and the reduced set [10]. From a biological perspective, genetic distance is defined as the accumulated differences in alleles at each locus [20]. This was calculated based on the allelic frequencies observed from a given set of microsatellite markers using Equation (1). The genetic distance matrix was generated using the dist function implemented within the adegenet package in R version 4.2.2 [21].

$$D\left(a,b\right)=-\mathrm{l}\mathrm{n}\left(\frac{{\sum }_{k=1}^v{\sum }_{j=1}^{m\left(k\right)}{p}_{aj}^k{p}_{bj}^k}{\sqrt{{\sum }_{k=1}^v{\sum }_{j=1}^{m\left(k\right)}{\left(p_{aj}^k\right)}^2}\sqrt{{\sum }_{k=1}^v{\sum }_{j=1}^{m\left(k\right)}{\left(p_{bj}^k\right)}^2}}\right)$$

(1)

In this study, a marker selection algorithm was developed to effectively decrease the number of microsatellite markers used in population genetic studies. This was achieved by enhancing the ACO algorithm for marker selection [22] and utilizing PIC as an informative marker indicator [17,23]. The PIC for each microsatellite marker was calculated using the PopGenUtils package in R version 4.2.2 [21]. In the microsatellite selection scheme, loci were sorted based on their PIC and the highest-ranking microsatellite was integrated into the selected marker set.

2.2. Ant Colony Optimization Algorithm

The ACO algorithm was used to select an optimal set of microsatellite markers. The ACO algorithm, inspired by the natural behavior of ants, is a metaheuristic optimization technique [7]. To facilitate the application of the ACO algorithm, the search space was represented by a directed graph [24] with 2 × N nodes, where N denotes the total number of microsatellite loci [8]. The ant pheromones were randomly distributed along the pathways. During each iteration, the ants independently construct their solutions by probabilistically selecting pathways based on pheromone trails, which serve as indicators of the solution quality. Once all the ants have constructed their solutions, the pathways are sorted based on their quality, and the corresponding pheromone trails are updated. The ACO algorithm was then executed with the appropriate parameters to identify discriminant microsatellite loci (

Table 1. Parameter used for the ant colony optimization algorithm [7,8].

Parameter	Description	Value
ant_n	Ant population size	50
E	Number of epochs (iterations)	120
α 1	Weight factor of the pheromone trail in the decision-making process	0.7
decay 2	Evaporation rate of the pheromone trail	0.9

¹ A higher value of α increases the significance of the pheromone trail, making the ants more likely to choose edges with stronger pheromone concentrations. ² A small value of decay allows the avoidance of becoming stuck on local minima and the encouragement of ants to explore new pathways.

). Finally, the initial pheromone values were adjusted based on the PIC of each microsatellite marker. Microsatellites with high levels of polymorphisms were preferred to those with low levels. This approach aims to reduce the computational noise, minimize the number of required iterations, and avoid potential entrapment in local solutions [25]. The described panel optimization algorithms were implemented using a Python version 3.11 [26] script (File S1) and executed on a Linux Ubuntu server version 18.04 [27].

2.3. Microsatellite Marker Dataset

The microsatellite selection scheme was evaluated using two datasets obtained from genetic diversity studies: a chicken genotyping dataset and a Chinese goral genotype dataset. The chicken dataset, from the Siam Chicken Bioresource Consortium Project, encompassed 652 individuals, was analyzed using 28 marker loci and available from https://doi.org/10.5061/dryad.hhmgqnkm0 (accessed on 5 July 2023) [28,29,30,31]. The genotype information of 79 individuals across 11 markers in the Chinese goral dataset was downloaded from https://doi.org/10.5061/dryad.wstqjq2hm (accessed on 5 July 2023) [32,33]. The datasets used in this study were formatted using the GenAlEx tool version 6.51 [34] and were compatible with Microsoft Excel. The number of alleles per locus (N_a), effective number of alleles (N_ea), observed and expected heterozygosities (H_o and H_e), and allele richness (AR) were evaluated for each microsatellite locus in both datasets. The PIC was computed using the “PIC” function available in the polysat package within R version 4.2.2 [35].

2.4. Comparative Evaluation of Marker Selection Schemes: ACO Algorithm, PIC, PIC + ACO, and Random Selection

A microsatellite marker selection model was fitted to minimize the loss of AGD accuracy. Four marker-sampling methods were used in this study. The first method employed in this study was the use of the ACO algorithm to select the most accurate panel without prior information regarding the polymorphisms of each locus. The second method involved sorting microsatellites based solely on their PIC and selecting the most informative loci. The third method involves ranking microsatellites based on their PIC and subsequently optimizing the set using PIC + ACO. A random selection scheme was used for the control group. Pairwise comparisons between selection schemes were conducted using the Tukey honest significance test, using the “pairwise_tukeyhsd” function from the statsmodel package [26]. The performance of each selection scheme was assessed through statistical pairwise comparisons using Tukey’s honest significance test. This analysis was conducted using the “pairwise_tukey_hsd” function from the statsmodel package in Python version 3.11 [26]. The PIC + ACO algorithm was used to progressively reduce the number of microsatellite markers to N = 2. The accuracy losses of the estimated values for H_o, H_e, and AR were evaluated. The AGD was reported, and graphical illustrations were generated using the “boxplot” function from the matplotlib package in Python version 3.11 [36]. Statistical regression analysis was conducted using the “OLS” function from the statsmodel package [37]. The estimation accuracy loss of H_o and H_e was determined by gradually reducing the number of microsatellite markers using the “plot” function from the matplotlib package in Python version 3.11 [36].

2.5. Estimation of Genetic Diversity Measurement on a Reduced Set of Microsatellite Markers

The microsatellite marker panel was assessed for each dataset by setting arbitrary error tolerances to 1%, 5%, and 10%. As a result, three reduced marker panels were created for chicken: GGA₁ (1% error tolerance-reduced marker), GGA₅ (5% error), and GGA₁₀ (10% error), and three marker panels for Chinese goral: NGR₁ (1% error), NGR₅ (5% error), and NGR₁₀ (10% error). The N_a, N_ea, AR, and PIC of the given population were evaluated in all microsatellite datasets, focusing on two statistical aspects: the mean difference between the measurements on the optimized and full sets, and the significance of the association of a higher measurement with the optimized set. The mean difference was used to explain the extent of deviation between the values reported for the full and reduced sets of microsatellites. The statistical p-value was calculated using an independent t-test and classified into four levels of significance: not significant (p > 0.05), slightly significant (0.01 < p < 0.05), moderately significant (0.001 < p < 0.01), and highly significant (p < 0.001). The statistical test was performed using the “ttest_ind” function from the stats package in Python version 3.11 [38]. The results were subsequently visualized using the “boxplot” function from the matplotlib package in Python version 3.11 [37]. The impact of reducing the number of microsatellites in a marker panel on population structure estimation was studied using three analytical methods: the Bayesian clustering algorithm [39], phylogenetic relationship analysis [40], and multidimensional scaling [41]. Population clustering analysis was conducted using Structure software version 2.3.4 [42]. The appropriate number of population clusters was determined by selecting the highest value of the Delta-K statistic, following the guidelines provided in the STRUCTURE software user manual [43]. The genetic distance between subpopulations was computed for the phylogenetic analysis using the “hclust” function from the stats package in R version 4.2.2 [35]. The dimensional scaling analysis was conducted using both principal component analysis (PCA) [44] with the “cmdscale” function from the stats package in R version 4.2.2 [35] and the discriminant analysis of principal components (DAPC). The resulting dimensional coordinates were visualized using the “dapc” function from the adegenet package in R version 4.2.2.

3. Results

3.1. Pairwise Comparison of Marker Selection Schemes on Two Genotype Datasets

The chicken and Chinese goral genotype datasets comprise N_a ranging from 5 to 82 alleles (average: 21), N_ea spanning from 1.14 to 26.22 (average: 6.40), AR ranging from 0.01 to 0.16 (average: 0.06), and PIC values ranging from 0.12 to 0.95 (average: 0.70) (Table S1). A comparison of the three selection methods indicated that the PIC + ACO selection scheme demonstrated superior accuracy on the chicken dataset for all marker quantities (N), except for N = 5 and N = 4, which showed statistical significance (p < 0.01). However, the ACO selection scheme was the most accurate for N = 5, whereas the PIC selection method showed the highest accuracy for N = 4. By contrast, for the Chinese goral dataset, the PIC + ACO scheme was the most accurate for marker sets consisting of nine, seven, and four loci. The highest accuracy was observed for marker sets comprising ten and eight microsatellites in the ACO scheme. However, for other values of N, higher accuracy was observed with randomly selected microsatellite markers than with the ACO, PIC, and PIC + ACO selection schemes (Tables S3 and S4; Figure S1).

3.2. Microsatellite Panel Selection Using Error Margins of 1%, 5%, and 10%

In the chicken dataset, with an error margin of 1%, the PIC + ACO selection method identified two microsatellites (LEI0094 and MCW0123) that could be excluded. Similarly, the ACO and PIC selection schemes each identified one microsatellite (MCW0206 and ADL0278, respectively) that could be excluded. With a permitted AGD estimation accuracy loss of 5%, the PIC + ACO selection scheme indicated the need for 12 marker loci. Based on the PIC selection policy, 13 markers were considered effective. The ACO selection algorithm required 13 markers, with 7 markers (MCW0034, MCW0183, LEI0192, MCW0123, LEI0234, MCW0069, and MCW0111) commonly selected by all three methods, including the ACO, PIC, and PIC + ACO selection schemes. Considering a threshold of 10% for AGD measurement, all three selection methods indicated the usability of 7 microsatellite markers, with 4 markers (LEI0234, MCW0104, LEI0192, and MCW0111) commonly selected by both methods. In the Chinese goral dataset, considering a 1% error allowance, all selection methods indicated that a full set of 11 markers was necessary. By selecting an error margin, the same set of markers consisting of 10 microsatellite markers, excluding SY259F, was reported by both the PIC and ACO selection schemes. In total, 9 microsatellite markers were identified as usable using the PIC + ACO selection method, excluding SY259F and SY128F. With an error margin of 10%, the ACO selection method determined that 8 microsatellite markers were adequate, excluding SY259F, SY76F, and SY449F. By contrast, the same set of 6 microsatellite markers (SY434F, SY14F, SY12BF, SY129F, SY449F, and SY128F) were identified using both the PIC and PIC + ACO selection schemes (Figure 1;

Table 2. Microsatellite marker panel selected by the 3-selection scheme using different accuracy loss margins.

Dataset	Average Genetic Distance Estimation Accuracy Loss		Selection Scheme
		PIC + ACO 1	ACO 2	PIC 3
Gallus gallus 28 markers	10%	MCW0034, MCW0104, LEI0234, MCW0016, MCW0111, MCW0183, LEI0192	MCW0104, LEI0234, LEI0166, MCW0123, MCW0111, ADL0268, LEI0192	MCW0034, MCW0104, LEI0234, MCW0123, MCW0111, LEI0094, LEI0192
	5%	MCW0034, MCW0104, MCW0165, LEI0234, MCW0123, MCW0206, MCW0111, LEI0094, MCW0183, MCW0069, LEI0166, LEI0192	MCW0034, MCW0078, MCW0098, MCW0165, LEI0234, MCW0216, MCW0123, MCW0206, MCW0111, MCW0183, MCW0069, ADL0268, LEI0192	MCW0034, MCW0104, MCW0330, LEI0234, MCW0123, MCW0016, MCW0111, LEI0094, MCW0183, MCW0069, MCW0295, ADL0268, LEI0192
	1%	MCW0034, MCW0098, MCW0081, MCW0330, MCW0165, LEI0234, MCW0222, MCW0206, MCW0104, MCW0078, ADL0112, MCW0216, MCW0111, MCW0183, MCW0069, ADL0268, LEI0192, MCW0037, MCW0248, MCW0014, MCW0103, MCW0067, MCW0016, MCW0295, LEI0166, ADL0278	MCW0034, MCW0098, MCW0081, MCW0330, MCW0165, LEI0234, MCW0222, MCW0104, MCW0078, ADL0112, MCW0216, MCW0111, MCW0183, MCW0069, ADL0268, LEI0192, MCW0037, MCW0248, MCW0014, LEI0094, MCW0103, MCW0067, MCW0123, MCW0016, MCW0295, LEI0166, ADL0278	MCW0034, MCW0098, MCW0081, MCW0330, MCW0165, LEI0234, MCW0222, MCW0206, MCW0104, MCW0078, ADL0112, MCW0216, MCW0111, MCW0183, MCW0069, ADL0268, LEI0192, MCW0037, MCW0248, MCW0014, LEI0094, MCW0103, MCW0067, MCW0123, MCW0016, MCW0295, LEI0166
Naemorhedus griseus 11 markers	10%	SY434F, SY14F, SY12BF, SY129F, SY449F, SY128F	SY434F, SY14F, SY12BF, SY129F, SY449F, SY128F	SY434F, SY14F, SY12BF, SY93F, SY129F, SY128F, SY84BF, SY84F
	5%	SY434F, SY14F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY84BF, SY84F	SY434F, SY14F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY128F, SY84BF, SY84F	SY434F, SY14F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY128F, SY84BF, SY84F
	1%	SY434F, SY14F, SY259F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY128F, SY84BF, SY84F	SY434F, SY14F, SY259F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY128F, SY84BF, SY84F	SY434F, SY14F, SY259F, SY12BF, SY93F, SY129F, SY76F, SY449F, SY128F, SY84BF, SY84F

¹ PIC + ACO, selection scheme involving ranking the markers by their polymorphic information content and subsequently optimizing the set using the PIC + ACO algorithm. ² ACO, selection scheme using only the ant colony optimization algorithm without any prior information on the PIC of the markers. ³ PIC, selection scheme sorting microsatellites on their PIC and selecting the most informative loci.

).

3.3. Genetic Diversity Expressed by the Reduced Set of Microsatellites Using Error Margins of 1% (GGA₁ and NGR₁), 5% (GGA₅ and NGR₅), and 10% (GGA₁₀ and NGR₁₀)

Biased values of genetic diversity were observed between the full and reduced sets of microsatellites when employing the aforementioned markers, with varying levels of statistical significance and discrepancy. On the chicken dataset, the highest divergence in N_a was observed on the reduced set of microsatellites, which had an average of 26.88 alleles (1.02-fold higher than the full set of loci), 37.83 alleles (1.44-fold), and 48.14 alleles (1.83-fold) with the GGA₁, GGA₅, and GGA₁₀ marker sets, respectively. Higher values of N_ea were observed on the GGA₅ and GGA₁₀ marker sets, with 10.97 (1.38-fold) and 12.6 (1.58-fold), respectively, whereas a negative discrepancy was observed in the GGA₁ marker set, with an average N_ea of 7.49 (0.94-fold). Similarly, the GGA₁ exhibited negative discrepancy in N_ea, AR, PIC, H_o, and H_e: the measured AR was 0.04 (0.98-fold), PIC was 0.75 (0.95-fold), H_o was 0.59 (0.98-fold) and H_e was 0.82 (0.99-fold). Conversely, the GGA₅ and GGA₁₀ yielded relatively high values: their AR values were 0.06 (1.4-fold) and 0.08 (1.79-fold); their reported PIC 0.86 (1.07-fold) and 0.88 (1.12-fold); the determined H_o 0.66 (1.10-fold) and 0.68 (1.13-fold); and the H_e 0.88 (1.06-fold) and 0.90 (1.08-fold), respectively.

For the Chinese goral dataset, discrepancy analysis could only be performed for the NGR₅ and NGR₁₀ microsatellite sets because the NGR₁ was not a reduced marker panel. The N_a allele exhibited an average of 8.66 alleles (1.01-fold) for NGR₅ and 9.33 alleles (1.09-fold) for NGR₁₀. The N_ea averaged a value of 2.27 (0.94-fold) for NGR₅ and 2.86 (1.19-fold) for NGR₁₀. The AR averaged a value of 0.11 (1.01-fold) for NGR₅ and 0.11 (1.09-fold) for NGR₁₀. The PIC yielded an average value of 0.46 (1.01-fold) for NGR₅ and 0.52 (1.14-fold) for NGR₁₀. H_o averaged a value of 0.16 (0.87-fold) for NGR₅ and 0.22 (1.21-fold) for NGR₁₀. The H_e yielded an average value of 0.48 (1.01-fold) for NGR₅ and 0.54 (1.13-fold) for NGR₁₀ (Figure 2; Table S2).

Previously described values were used to demonstrate the correlation between microsatellite panel quality and population genetic measurements at different levels of significance. In the GGA₅ marker panel, moderately significant associations (p < 0.01) were observed for N_a, N_ea, and AR, and low statistical significance (0.01 < p < 0.05) was determined for PIC, H_o, and H_e. For GGA₁₀, N_a and AR were determined to have high statistical significance (p < 0.001), N_ea exhibited moderate statistical significance (0.001 < p < 0.01), PIC and H_e had low statistical significance (0.01 < p < 0.05), and H_o had no statistical significance. However, for the chicken GGA₁ and Chinese goral datasets (NGR₁, NGR₅, and NGR₁₀), insufficient data used for the statistical tests hindered the achievement of statistically significant findings (

Table 3. Statistical significance of the association of the number of alleles (N_a), the number of effective alleles (N_ea), the allele richness (AR), the polymorphic information content (PIC), the observed (H_o), and the expected heterozygosity (H_e) with the reduced microsatellite marker panel.

Dataset	Reduced Panel	Measurement	Mean-Diff	t-Stat	p-Val	Significance
Gallus gallus 28 markers	GGA1 (26 markers)	Na	5.115	−0.394	0.697	ns
		Nea	6.813	−1.909	0.067	ns
		AR	0.008	−0.397	0.695	ns
		PIC	0.122	−1.341	0.192	ns
		Ho	0.101	1.975	0.108	ns
		He	0.099	2.354	0.193	ns
	GGA5 (12 markers)	Na	18.521	3.240	0.003	**
		Nea	5.246	3.093	0.005	**
		AR	0.030	3.146	0.004	**
		PIC	0.110	2.515	0.018	*
		Ho	0.105	2.422	0.023	*
		He	0.086	2.347	0.027	*
	GGA10 (7 markers)	Na	27.857	5.081	0.000	***
		Nea	6.175	3.222	0.003	**
		AR	0.045	4.866	0.000	***
		PIC	0.129	2.586	0.016	*
		Ho	0.101	1.975	0.059	ns
		He	0.099	2.354	0.026	*
Naemorhedus griseus 11 markers	NGR1 (11 markers)	Na	–	–	–	–
		Nea	–	–	–	–
		AR	–	–	–	–
		PIC	–	–	–	–
		Ho	–	–	–	–
		He	–	–	–	–
	NGR5 (9 markers)	Na	0.667	0.251	0.808	ns
		Nea	0.668	−0.595	0.567	ns
		AR	0.008	0.228	0.825	ns
		PIC	0.015	0.087	0.933	ns
		Ho	0.130	−0.899	0.392	ns
		He	0.026	0.147	0.886	ns
	NGR10 (6 markers)	Na	1.733	0.874	0.405	ns
		Nea	1.022	1.249	0.243	ns
		AR	0.023	0.892	0.396	ns
		PIC	0.142	1.135	0.286	ns
		Ho	0.087	0.771	0.460	ns
		He	0.140	1.081	0.308	ns

ns: No significant association (p > 0.05). *: Weak significance association (0.05 < p < 0.01). **: Medium significance association (0.01 < p < 0.001). ***: High significance association (p < 0.01).

).

3.4. Comparison of Population Structure Inference between the Full Set and Reduced Sets of Microsatellites

The presence of two population clusters (K = 2) was revealed in the downstream analysis of the chicken population genotype dataset using STRUCTURE software. Regardless of the number of microsatellite markers used for the population genetics assessment, the same value of K = 2 was consistently observed (Table S4; Figure S2). Visualization of population genetics and microsatellite marker panel accuracy can be achieved using STRUCTURE, phylogenetic trees, PCA, and DAPC plots (Figure 3, Figures S3 and S4). All 31 chicken subpopulations were classified into K = 2 clusters with statistical significance for the posterior probability (p < 0.01) for the four studied marker panels (GGA₁, GGA₅, GGA₁₀, and the full set of 28 chicken microsatellites). For K = 7, 28 of the 31 subpopulations were successfully clustered into 7 groups using the full set of 28 microsatellites with statistical significance (p < 0.01). With GGA₁, the number of clustered subpopulations remained at 28, whereas GGA₅ clustered 29 subpopulations and GGA₁₀ 26 subpopulations. For K = 9, 30 out of 31 subpopulations were assigned to 9 clusters using the full set of 28 markers, whereas both the GGA₁, GGA₅, and GGA₁₀ marker panels reported 29 clustered subpopulations (Figure 3; Table S5). However, with the use of a reduced set of microsatellite markers, different values were reported, and no inferred clusters were revealed in the membership probability structure, PCA, and DAPC analysis. Because there was only one genetic subpopulation in the Chinese goral dataset, no statistical comparison of subpopulation clustering could be inferred.

4. Discussion

Genetic researchers face the challenge of an increasing number of usable microsatellite panels, prompting the need for smart and efficient selection of markers in the fields of genetic diversity, population genetics, and breeding programs. A trade-off between cost and result quality must be made, considering research expenses and time as limiting factors. In previous studies, various marker selection algorithms have been investigated, including the k-optimal [45], decision-tree induction algorithm [46], traveling salesman [13], ant colony algorithm [8], and genetic algorithm [11]. Considering panel selection as an optimization problem, any of the previously studied algorithms can be used as they offer a cost function to minimize or maximize [16].

4.1. Challenges in Microsatellite Marker Panel Selection

The informativeness of microsatellite markers is directly related to their degree of polymorphism [17]. The polymorphism exhibited by each marker (locus) should be considered when constructing a microsatellite panel [47]. A reduced panel of 9–12 markers was considered suitable. However, in genetic diversity and population analyses of species such as chickens, cattle, and dogs, the use of 18–30 markers is common. These species, which are known for their numerous varieties and breeds, have been studied and improved through breeding programs using microsatellite standard sets. However, considerable variations have been observed in the effectiveness and accuracy of each available microsatellite marker panel. The quality of the results is largely dependent on the choice of the marker set, as not all microsatellite panels are equivalent [48,49]. Usable and convenient microsatellite markers can be identified by combing through past studies; however, a universal optimized marker panel does not exist because of the varying genetic marker specifications across different research domains [50,51]. Another method uses the PIC, allele variation (N_a/N_e), AR and H_e as informativeness indicators of a particular locus [49,52]. The use of a well-selected panel could also compensate for certain genotyping errors and estimate population genetic measurements within an acceptable accuracy loss [10,53].

The PIC has always been regarded as an accurate quality indicator of microsatellite markers; however, the developed selection scheme does not prioritize the highest PIC microsatellites [17,23]. With the chicken dataset, of the reported 7-microsatellite set, LEI0094 and MCW0123, despite having high PIC values—0.93 and 0.88—respectively, were excluded. Instead, our marker selection scheme (PIC + ACO) included MCW0183 and MCW0016, which have PIC values, of 0.83 and 0.87, respectively. Similarly, among the 14 microsatellite marker sets, MCW0016, MCW0295, MCW0330, and ADL0268 (with PIC values of 0.87, 0.84, 0.85, and 0.85, respectively) were excluded, whereas LEI0166, MCW0165, and MCW0206 (with PIC values of 0.74, 0.69, and 0.81, respectively) were selected. This suggests that the accuracy of individual identification is not always guaranteed by the highest PIC markers, as microsatellite markers can provide redundant information due to non-random associations between distant loci [54]. However, regardless of the chosen accuracy loss threshold, all markers with low PIC values are generally excluded by the PIC + ACO selection scheme, with an allowed accuracy loss of 10%, all markers with PIC lower than 0.83 are excluded, and a loss tolerance of 5% excludes all markers with PIC below 0.69. This suggests that PIC provides valuable insights into the efficiency of molecular markers for genetic studies, as stipulated by Serrote et al. [17]. Publicly available microsatellite panels for genetic studies and chicken breeding programs are generally highly polymorphic [5,28,29,30,31]. Similarly, in the second dataset, the same set of markers was reported using the PIC and PIC + ACO selection schemes for margin tolerances of 1% and 10%, respectively. However, with a 5% margin tolerance, PIC + ACO excluded SY128F, which was among the top two highest PIC microsatellites in the dataset. In addition, the highest PIC markers were always selected by the PIC + ACO method for 1% and 10% error tolerances. Referring to the chicken dataset used in this study, an average genetic distance accuracy loss ranging from 5% (GGA₅) to 10% (GGA₁₀) was observed. The chicken genotype dataset revealed that the 7 most informative microsatellites were MCW0111, LEI0234, MCW0034, MCW0016, LEI0192, MCW0183, and MCW0104 markers. These markers exhibited higher effectiveness (PIC > 0.83, N_a > 28, N_ea > 6.79, H_o > 0.58, and H_e > 0.85), as suggested by previous studies on chicken population genetics [30,55]. Moreover, the clustering of the putative chicken population was accurately displayed by visual representations of PCA and DAPC using the 7 selected markers mentioned above. Microsatellite marker set reduction could be further pursued by increasing the accuracy loss margin by up to 15%, as reported by Xiong et al. [54] for other types of molecular markers. The relevance of the proposed microsatellite panel size was further supported by experiments on the Chinese goral dataset, which did not yield any marker combination with fewer than 9 markers (NGR₅).

Microsatellite panels with high levels of genetic diversity are widely available for numerous species, therefore expanding the applicability and scope of this study [28,56]. The algorithm studied was well-suited for refining a large set of microsatellites (more than 20 microsatellite sets) with sufficient alleles to allow for some accuracy loss in the genetic measurement estimations. Using this algorithm, significant budgetary savings can be achieved by excluding a substantial number of microsatellite markers. Moreover, valuable insights into the efficiency of microsatellites and their individual contributions to the effectiveness of marker panels can be obtained [47]. However, the heterozygosity of individuals is not considered by the AGD function used to assess genetic diversity among populations [20], causing the algorithm to disregard valuable information on gene diversity and inbreeding within populations. Moreover, failures during microsatellite marker amplification and genotyping processes have been omitted in almost all studies [57], potentially leading to the exclusion of some usable microsatellite markers for population genetic investigation [58].

4.2. Using the PIC as a Discriminative Power Indicator of the Marker

The ant colony optimization (ACO) algorithm, which was proposed in the early 90s as an approach to resolving optimization problems, has garnered interest because of its simplicity and versatility [7]. It exists in numerous variants, including the ant system (AS), ant-Q, max-min ant system, rank-based ant system, BWAS, and hypercube AS [59,60,61,62]. The ACO algorithm, which belongs to the group of metaheuristic approaches [14], shares commonalities with trending optimization algorithms, such as the genetic algorithm (GA), particle swarm (PSO), or seagull optimization algorithm (SOA). It determines the optimal solution by spreading pheromones on pathways based on the solution quality [8]. Properly balancing exploration and exploitation in the algorithm parameters is crucial to avoid infinite loops or becoming stuck in local solutions [7]. Similar to the trial and reward concept used in reinforcement learning, every possibility of the microsatellite panel was assessed using the optimization pipeline used in the ant colony optimization algorithm, and a quality score was assigned to each based on certain criteria [63]. The original version of the ant colony optimization algorithm formulated by Colorni et al. [7] used a stochastically generated initial solution that was gradually improved. However, the discriminative power of markers is closely related to various variables, including N_a, N_ea, AR, and PIC [17,20]. This led to the investigation of a method that includes this information as an initial variable to be progressively improved by the heuristic algorithm. For the chicken dataset, a comparative study of the four selection schemes revealed that the accuracy of the improved algorithm (PIC + ACO scheme) was higher than that of the original algorithm (ACO). With the optimized chicken microsatellite and 5% accuracy loss, 3 highly polymorphic markers (MCW0104, LEI0094, and LEI0166) were omitted by ACO but included in the GGA₅ panel.

4.3. Implications for Conservation Effort and Breeding Program

The chicken and Chinese goral datasets used in this study were sufficiently large to facilitate the use of the marker optimization algorithm [28,29,30,31,32,33]. The availability of a large genotype dataset allows for a more optimized exploration of the marker efficiency mechanism. In addition to the widely developed non-invasive sampling methods [64], the assessment and elucidation of genetic diversity can be significantly enhanced by the development of molecular markers. Population dynamics and migration in several animals have been studied using non-invasive fecal sampling [65]. However, the quality of the DNA stock after extraction is very low, and not all common sets of microsatellite genotyping are applicable. The competency of the output results in the full set can be effectively predicted by optimizing the microsatellite marker panel. Conservation and breeding initiatives can be greatly enhanced by the in silico development of microsatellite markers, enabling a more optimized fit for the proposed microsatellite panel reduction scheme presented in this study [66]. Budgetary barriers to numerous conservation and breeding initiatives would be considerably alleviated by this approach, offering an opportunity for population monitoring within an acceptable accuracy loss in conservation and breeding programs. Interestingly, the number of markers that can be amplified in a single reaction significantly influences both cost and efficiency. This relationship offers opportunities for cost reduction. Although marker multiplexing effectively manages this trade-off, PCR efficiency is not closely tied to polymorphism. In our current study, we prioritize polymorphism, leaving the amplification efficiency of markers as a potential focus for future research.

5. Conclusions

This study explored the use of a modified ACO algorithm, PIC + ACO selection scheme, to determine the most effective microsatellite panel for genetic diversity research with different accuracy loss tolerances. Experiments on both datasets revealed that microsatellite markers allow for the exclusion of many markers while maintaining acceptable precision in population genetics assessment. The optimized reduced set of markers exhibited efficiency related to various metrics. However, the PIC + ACO selection scheme shows that markers rely on hidden variables beyond simple metrics. The study results show that reducing laboratory costs could promote conservation initiatives and population genetic investigations in biodiversity conservation and breeding programs for genetic improvement.