The possibility to predict haplotypes and, therefore, alleles of ungenotyped variant loci arose from the fundamental concept in genetics that individuals of a given population share the haplotype stretches that have descended from common ancestors. This is evident when investigating the transmission of alleles into families (Malerba et al., 2000). Siblings share a higher number of identical stretches of DNA, deriving from parents, than a randomly chosen group of individuals. Therefore, individuals having a common ancestor are expected to share haplotype stretches whose extent depends on the temporal distance (number of recombination events throughout generations) between the individuals and the common ancestor.

The idea that individuals share identical genetic stretches (defined as haplotypes) led researchers to explore and implement imputation in genetic studies: genetic similarities among individuals suggested they could share the same ancestral haplotype and hence this turned out to be effective in genotype reconstruction of missing genotypes in related individuals. Since imputation is based on the linkage disequilibrium structure of a given genomic region, it is important to phase the alleles at different close variant loci which is the process of addressing marker alleles to either maternal or paternal haplotypes. This task can be challenging in unrelated individuals whereas it could be easier when performing family-based studies to infer genotype distribution in pedigrees and test genetic models of inheritance for traits and diseases (George and Elston, 1987; Fulker et al., 1999).

The general procedure behind imputation, originally developed on pedigrees, was based on three main steps: (1) some samples belonging to the same family were genotyped on several thousands (or hundreds of thousands) of markers distributed across the genome, (2) information on the haplotype to which these markers belonged were arranged and collected and (3) untyped markers of the remaining samples (of the same family) were inferred, recognizing the shared haplotypes with the fully characterized individuals (Fig. 1). The whole process was entitled “in silico genotyping” (Burdick et al., 2006), most likely because of its ability to predict genotypes from missing genetic information (Fig. 1).

Sooner, research interest in imputation moved to samples of unrelated individuals, which mostly build up the core of case-control studies. Compared to pedigrees, unrelated individuals did not share long haplotypes (as in the case of family members), but shorter regions that, in any case, could still be identical-by-descent, because of the presence of common ancestors, albeit far distant in time. Because a direct, informative comparison was not possible, as in the case of consanguineous individuals, the imputation process needed to be slightly revised from its original design. Specifically, (1) a catalog of detailed haplotypes for the genotyping of many individuals was set (reference panel); (2) then, phasing of close markers and comparison of phased haplotypes with the detailed haplotypes of the reference panel was performed, and (3) finally missing genotypes were predicted based on the haplotypes of the reference panel matching the phased haplotypes of the sample set.

The analysis of unrelated samples started moving the focus of imputation from close little groups of related samples, at most made of tens of individuals (Marchani et al., 2012; Chen et al., 2013), to higher amounts of individuals, not necessarily related, and sharing at a certain point similar genetic information, the so-called reference panel (Browning and Browning, 2016). The modern imputation finally began. Today, in the big data and next-generation sequencing era, genotype imputation relies on enormous reference panels, made of hundreds of thousands of samples and millions of genome-wide markers (McCarthy et al., 2016). Since the reference panels derive from individuals belonging to many ethnic groups, imputation may be conducted using the proper reference panels, even though it is not still clear what is their most effective structure (for example, haplotypes from the world-wide population or from the population having the same ethnicity of the tested samples) (Roshyara et al., 2016; Degenhardt et al., 2019; Kowalski et al., 2019). To handle these data, fast and efficient algorithms, as well as powerful and performant computational resources, are constantly improved, to keep up with the pace of present-day needs.

The sample set is commonly made of tens or hundreds of individuals who have been genotyped for some thousands of markers along the whole genome. The reference panel, instead, is a set of several tens or hundreds of thousands of individuals who have been genotyped at the genomic level for several hundreds of thousands or millions of loci. In the reference panel, the genetic information of all the markers is well characterized and, for this reason, it acts as the template for comparisons with the sample set. Indeed, the typed markers in the sample set are matched with the markers stored in the reference panel; the aim of this comparison is to find matches between groups of markers, the haplotypes. Hence, the matching markers are used as seeds to reconstruct in every sample of the sample set the genetic information (or genotypes) of the surrounding markers.

Following the proposed workflow, imputation can infer intervals of imputed variants surrounded by stretches of genotyped markers; imputed variants are generated from the comparison between the sample set and the reference panel. However, this inference does not generate results having absolute certainty: imputed markers show a variable level of uncertainty, depending on several factors (see section Factors affecting imputation), and for this reason, cannot be represented as standard genotypes. Indeed, each imputed marker is reported not as a discrete genotype but as a genotype probability (GP), that is to say giving two alleles 0 and 1, the probability of an individual in the sample set of being homozygous for the reference allele (0/0), heterozygous (0/1), and homozygous for the alternative allele (1/1). To summarize these probabilistic values, imputed data are mostly represented in terms of allelic dosage (DS), a value between 0 and 2; allelic dosage, or dosage, reports for every marker the number of alternative (or risk) alleles carried by a single individual in the sample set. Usually, a dosage equal to 0 corresponds to a probability of 1 of being homozygous for the reference allele, a dosage equal to 1 to a probability of 1 of being heterozygous and a dosage of 2 of being homozygous for the alternative alleles. Because the allelic dosage is a continuous value, all the possible values between 0 and 2 represent a probability of being closer to one of the three playing genotypes.

The first software implementations were extremely intensive and required a huge amount of computational power, memory, and data storage. During the past years, the methods underwent considerable improvements, becoming faster, more efficient, and more effective in every step of the genotype imputation procedure. Today, software implementation allows to impute not only bi-allelic SNPs but also multi-allelic variants (either SNPs or small INDELs); recent studies pointed out that all previously developed tools performed well on bi-allelic variants, but only new generation tools can handle multi-allelic variants, producing reliable results (Hanks et al., 2022). Indeed, current genotype imputation software is mostly categorized into two significantly different methods: population-based imputation (PBI) and family-based imputation (FBI) (Liu et al., 2019). The PBI methods are the most used and well-established approaches for genotype imputation; they usually rely on a reference panel of unrelated samples and make use of linkage disequilibrium and correlation between close SNPs of a specific population to predict ungenotyped markers. On the other hand, FBI methods are used to infer genotypes of related individuals; they rely on familial information, such as pedigrees and identity by descent, to reconstruct allelic phases and infer unobserved genotypes (Saad and Wijsman, 2014a). Different studies investigated the strengths and weaknesses of these methods (Saad and Wijsman, 2014b; Liu et al., 2019; Ullah et al., 2019). PBI showed to be particularly suitable for the investigation of common variants (frequency greater than 5%), having a higher imputation accuracy when the investigated sample set is closely related to the population of the reference panel; moreover, the size of the reference panel showed to increase imputation accuracy for rarer variants (frequency lower than 5%), probably due to their higher frequencies in a larger dataset. On the contrary, FBI methods are widely used for the imputation of rare variants, due to their increased frequencies in small and inbred populations, such as in families, but seemed to perform poorly on common variants; however, the use of study-specific reference panel seems to increase imputation accuracy for common variants, probably due to higher precision in the haplotypes of local reference samples (Liu et al., 2019; Whalen et al., 2019). For all these reasons, hybrid approaches have been explored (Liu et al., 2013; Kreiner-Møller et al., 2015; Lent et al., 2016), to make good use of the advantages of the two strategies, showing a considerable increase in genotype imputation performance (Ullah et al., 2019). Of note, nowadays most used and standard software are Beagle 5.4 (Browning et al., 2018, 2021), Eagle 2.4.1 (Loh et al., 2016, 2016) and SHAPEIT4 (Delaneau et al., 2019) for phasing and Beagle 5.4 (Browning et al., 2018, 2021), IMPUTE5 (Rubinacci et al., 2020) and Minimac4 (Das et al., 2016) for population-based imputation, and GIGI (Cheung et al., 2013), Merlin (Burdick et al., 2006) and cnF2freq (Nettelblad, 2012) for family-based imputation. In the end, the most promising hybrid approaches, taking advantage of both PBI and FBI methods, are Fimpute (Sargolzaei et al., 2014), and the combination of IMPUTE2 with Merlin (Liu et al., 2019) and GIGI with Beagle (Saad and Wijsman, 2014a).

The constant increase of computational load, mostly related to the number of input markers and samples and the size of the reference panels, pointed out the need for powerful infrastructure to perform heavy computational tasks. Two free next-generation genotype imputation servers have been developed, respectively the Michigan Imputation Server and the TOPMed Imputation Server (Das et al., 2016). The platforms consist of a user-friendly interface in which the user, upon registration, can input their sample data (as a gzipped variant call format, one for each chromosome), setting the desired parameters like choosing standard reference panels (such as the 1000 Genomes Project phase 3, the Haplotype Reference Consortium panel and the TOPMed r2 panel; see section Reference panels) and an optional subpopulation, quality filtering to apply on the imputed data (see section Quality filtering for good imputation results), and finally choosing to perform phasing, imputation or both the two procedures sequentially. Both Michigan and the TOPMed Imputation Servers use Eagle 2.4 and Minimac4 to perform phasing and imputation, respectively. Although the combination of Eagle 2.4 and Minimac4 was assayed as one of the slowest approaches when compared to other software (such as Beagle 5), it turned out to be the most efficient in terms of computational resources for every size of the input sample set (Browning et al., 2018), representing the most suitable solution for online queued platforms.

In general, genotype imputation requires high amounts of time and resources. State-of-the-art tools and algorithms have boosted performances, not only in managing big data but also in lowering the running time and computational load. Moreover, computation time and sample size are indirectly related: per-sample running time to perform imputation increases together with the decrease in sample size. This counterintuitive procedure is related to the fixed computational cost of reading the reference panel, which is shared among all the samples (Browning et al., 2018). Thus, the imputation procedure focuses on groups rather than on single individuals; this overview is confirmed not only from a constitutive perspective (that is, the need to have a group of individuals in both the sample and the reference panels, to clarify as much as possible the frequencies of the haplotypes in the working dataset), but also from a computational perspective.

In the last couple of years, with the advent of cloud computing services (such as Amazon Web Services, Microsoft Azure, or Google Cloud), imputation pipelines started to be optimized on pre-defined and suited virtual machines, in order to dramatically reduce the cost of imputation: a recent study pointed out the possibility to perform imputation at a cost lower than 1 U.S. cent per sample (Browning et al., 2018). The fact that imputation has been implemented in the context of a global computation indicates how much it is a fundamental tool in the most current frontier research.

Historically, the first reference panels for genotype imputation came from international consortia: the International HapMap project and the 1000 Genomes Project. Both projects aimed to characterize the genetic variability of human populations.

The main differences among them were the number of samples and genetic markers available as well as the number of represented human populations and ethnicities. In chronological order, the first two panels were the HapMap phase I (International HapMap Consortium, 2005) and phase II (International HapMap Consortium et al., 2007) which comprised 269 samples across four populations (Yoruba in Ibadan, Japanese in Tokyo, Han Chinese in Beijing and U.S. Utah residence with northern and western Europe ancestry) on 1 and 3 million of SNPs, respectively. With the spreading of first- and second-generation sequencing technologies, the 1000 Genomes Project (1000 Genomes Project Consortium et al., 2010) came to light with its phase I (1000 Genomes Project Consortium et al., 2012) and phase III (1000 Genomes Project Consortium et al., 2015) panels, comprising 2504 individuals from 26 populations for over 88 million variants. In 2016, the 1000 Genomes Project established itself as the standard resource for genomic analysis, superseding the HapMap panels. However, in the last five years, alternative and more powerful resources have been developed: a clear example is the Trans-Omics for Precision Medicine (TOPMed) reference panel (Das et al., 2016), made of around 180 thousand sequenced individuals (60% of non-European ancestry) from more than 85 different studies for a total of 308 million variants (version r2).

One of the main features and concerns regarding the reference panels is the structure of populations and/or subpopulations included in the reference panel. The presence of a reference panel equipped with a reliable reference population, having a high number of markers and individuals (and hence, a good picture of the most common haplotypes across the population), is a fundamental step for good quality imputation results (see section Factors affecting imputation). During the comparison of samples of both unphased sample set and reference panel, genotype imputation can lead to stronger results when the two sets are highly similar (i.e., having a similar haplotype structure); on the contrary, when the genetic background of the sample and reference sets is quite distant, imputation is committed to a dramatical increasing of false and/or weak signals. As a rule, in order to increase the accuracy of imputation calls, a consensus between the ancestries of the sample and reference sets is fundamental (de Marino et al., 2022), but a suitable reference panel is not always available. Indeed, recent studies, which focused on sample sets of mixed ancestries or low-represented subpopulations, pointed out the limits of present-day reference panels. For this reason, several researchers (Kowalski et al., 2019; O’Connell et al., 2021) tried to manage this issue by setting reference panels including individuals from different populations to ideally mimic the ancestries of the mixed populations. As a result, careful study designs together with a combination of reference populations from different ancestries seemed to be one of the most feasible approaches to overcome the bias of low-represented populations.

Genotyping technology (arrays or sequencing) plays a crucial role in imputation performances. Since each genotyping platform spans differently along the genome, it also determines the different sets of markers that would be inferred by the imputation process. The technology impacts the number of variant sites that can be tested (Hanks et al., 2022). Moreover, not only the sample set but also the reference panels are affected by the genotyping techniques. Depending on how the reference panel has been generated (setup of variant sites included), imputation might achieve different results (i.e., only the variant sites present in the reference panel can undergo imputation). Panels of hundreds of thousands of markers widespread throughout the genome and representing different genomic locations (for example, regions having different genomic complexity) positively affect imputation, guaranteeing good-quality results. On the contrary, panels having low SNP density or a small population size would negatively impact imputation results (Das et al., 2018).

Imputation is also influenced by minor allele frequencies (the allele at a variant site having the lower frequency). Common variants (>5%), which are usually well represented across reference panels and present enough in sample sets, are smoother to impute than rarer variants. Moreover, imputation of low frequency (1%–5%) and rare (<1%) variants faces several limitations (Yu et al., 2022). Imputation can just infer known genetic information, which is stored in the reference panel; low-frequent to ultrarare (<0.1%) variants are usually not much represented in reference panels and, for this reason, would be inferred with low probabilities, resulting in unreliable calls (Zhang et al., 2021). Rare variants are in strict dependency on the chosen reference population: rare allelic frequencies are most of the time specific to a precise population, and the usage of a population not too close to the sample set may trigger a huge number of false positives (Hanks et al., 2022). However, several approaches have been developed to increase imputation accuracy for rarer variants, such as using imputation methods that combine population-based and family-based approaches (see section Present-day imputation approaches) and the usage of specific reference panels (see section Reference populations).

Despite the need for a reference panel, reference populations can backfire on imputed results. The choice of an accurate and sample-specific reference panel is crucial to ensure good quality results. However, a perfect match between sampled and reference individuals is only theoretically possible due to the large number of admixed samples in modern-day human populations (Hanks et al., 2022). Thus, the lower amount of matches between samples and reference set may increase the levels of uncertainty in imputed genotype calls, prominently for admixed populations. However, the application of state-of-the-art methodologies, together with a fine phasing of genetic data and a deeper analysis of the genetic background of samples, may mitigate population biases, generating reliable imputed results (de Marino et al., 2022). Several studies (Lin et al., 2018; Vergara et al., 2018; Bai et al., 2019) have investigated how imputation accuracy is affected by the presence of different human populations in the reference panel when analyzing underrepresented ethnic groups. The increasing distance between the sample set and the reference population negatively affects the imputation accuracy of common (>5%) and low-frequent variants (>1%); however, a small fraction of diversity in the reference panel seems to improve imputation performance in rare variants (<1%): examples have been reported for Han Chinese (Bai et al., 2019), Hawaiians (Lin et al., 2020), Turkish (Kars et al., 2021) and Latin Americans (Jiménez-Kaufmann et al., 2022). Nevertheless, genotype imputation of non-European samples still represents a challenging task, due to the predominance of European samples in most reference panels. To answer this need, several research projects started developing reference panels tailored to underrepresented populations: notable examples are the CAAPA reference panel for African Americans (O’Connell et al., 2021) (developed by the homonymous Consortium on Asthma among African-ancestry Populations in the Americas) (Mathias et al., 2016), the GAsP reference panel for Asian (GenomeAsia100K Consortium, 2019) (developed by the homonymous Genome Asia Pilot project together with the GenomeAsia 100K Project) and the ChinaMAP for Chinese (Li et al., 2021).

Other than the main factors affecting imputation quality, researchers identified finer and more specific parameters that need to be considered as a source of unreliable results: recombination rate, GC content, the distance between genotyped markers, the presence of structural variants and segmental duplications. Among these factors, a higher recombination rate, a lower GC content, the presence of structural variants, and segmental duplications have been found to be associated with lower-quality imputation results (Hanks et al., 2022), showing consistency and persistency across every method.

Before imputation, genotyped samples undergo several exclusion criteria (Li et al., 2009; de Marino et al., 2022; Hanks et al., 2022) to remove individuals having low call rates (<95%) and high missingness (>2%), significantly deviating from the Hardy-Weinberg Equilibrium (p-value < 10e-6) and duplicates. Moreover, allele labeling (reporting allele names referring to the same DNA strand for all the variant sites and all the sample sets, including the reference panel) between sampled markers and reference panel is checked, to avoid any comparison mistake, the harbinger of unreal mismatches or wrong strand reading (as for the markers where the two alleles are either A and T or C and G). In the end, samples are analyzed for their genetic background. Individuals are assayed through principal component analysis (PCA) for their genetic origins, to investigate the real ancestral background and to choose the most suitable reference panel for imputation.

To assay imputation quality and reliability, it is necessary to perform post-imputation quality control. Three main methods have been developed to assess imputation outcomes: concordance between genotypes, imputation quality score (IQS) and correlation between best-guessed genotypes (defined as R-squared or R2).

Assessing concordance between real and imputed genotypes represents a method to determine imputation quality. For every marker, genotype imputation infers a probabilistic value (in terms of genotype probabilities or alternative allele dosage), representing the probability of a genotype of being homozygous for the reference (0/0) or the alternative (1/1) allele or heterozygous (0/1). The approximation of genotype probabilities to a discrete genotypic value is far to be reliable and may lead to a higher level of wrong calls. Indeed, such an approximative method was previously tested (Abo et al., 2012) and benchmarked (Marchini and Howie, 2010) in early imputation studies, but was soon abandoned in favor of more functional metrics (de Marino et al., 2022).

Rather than approximating the probabilistic information to discrete genotypes, a more reliable way to evaluate imputation is to look at genotype probabilities directly. The first metric which was developed (Lin et al., 2010) was the imputation quality score (IQS). IQS discriminates between well-imputed and poorly imputed SNPs based on the genotype posterior probabilities. This metric compares the proportion of agreement between the haplotype-based and randomly imputed genotypes, weighted by the allelic frequencies of the imputed marker. Early usage of this metric showed remarkable improvement compared to the evaluation of the concordance approach (reported above), but since it is strongly dependent on allele frequency values, it is not suitable for rare variant sites.

Thus, the latest developed and most used metric is the R-squared (R2), defined as the correlation between the variance of the marker estimated from the counts of individual imputed alleles for every sample and the expected variance estimated from the overall allele frequencies. Hence, R2 is generally expressed as the ratio between the variance of the imputed alleles probabilities divided by the variance of the same alleles if they were perfectly imputed. In detail, R2 has been conceived in at least two different ways. Methods like MACH and Minimac express R2 values as the best approximation between the observed dosage variants on the expected dosage variants; instead, methods like Beagle and IMPUTE calculate the R2 values as the best approximation between the most likely genotype and the true unobserved expected allelic dosage (Chanda et al., 2012; Ramnarine et al., 2015).

For several years, GWASs based solely on genotyped samples which were at most assayed for tens of thousands of markers, a small portion if compared to our genome. The increasing interest in genotype imputation and the early work on related and unrelated individuals showed researchers that combining the discovery potential of GWAS together with the ability of imputation to infer genetic data from missing information would have led to an increase in the statistical power of genomic scans (Quick et al., 2020). Early evidence of this successful combination was pointed out in diseases, such as type 2 diabetes in the Finnish population (Scott et al., 2007) and familial cases affected by multiple sclerosis (Dyment et al., 2008), and in complex traits (Newton-Cheh et al., 2009), identifying novel loci explaining genetic susceptibility, risk, or sample variance.

The possibility of combining genetic information coming from genotyping techniques together with probabilistic inferred data, immediately revealed the strengths and weaknesses of this approach. On the one hand, enriching typed data for a great number of markers that have not been previously genotyped allowed to reach reliable association signals and to confirm previously reported evidence, as for psoriasis (Nair et al., 2009) but also to identify novel loci of understudied diseases, as for eosinophilic granulomatosis with polyangiitis (Lyons et al., 2019). However, on the other hand, the methods available to study genetic signals did not support the probabilistic perspective generated in the imputation process. Approximating genotype probabilities or allelic dosages to genotype was firmly discouraged (Li et al., 2009), and for this reason, novel methodologies had to be developed.

To overcome this initial bias in managing different representations of genetic data, several models that were able to account for probabilities and to handle uncertainty were implemented and integrated into association pipelines (Marchini and Howie, 2010). One of the first developed and probably most used models is the frequentist model. This model compares for every SNP a pre-determined model of association (such as additive, dominant, recessive, and others) against a model of no association, considering the genetic information as a likelihood probability. The frequentist model was immediately implemented since the earliest snptest release (Marchini et al., 2007; Wellcome Trust Case Control Consortium, 2007). However, the model faced immediate problems regarding spurious results, mostly due to small sample sizes, low allelic frequencies, and high levels of uncertainty in imputation calls (Lu et al., 2010). Novel approaches were developed to overcome these issues, mostly based on the Bayesian statistical method (Stephens and Balding, 2009). As for the frequentist model, Bayesian methods as well compare a model of association against a model of no association but implement a more complex and parametrized modeling. Despite the similarities and differences between these two approaches, genetic data which underwent imputation had to be considered as probabilistic, either in terms of genotype probabilities or allelic dosage, to ensure robust and unbiased association signals as demonstrated in several studies spanning from computational benchmarking (Song et al., 2018; Jørsboe and Albrechtsen, 2022) to genomic analyses investigating either disorder (Vissers et al., 2019) or traits (Tan et al., 2019).

The HLA complex presents the highest level of polymorphisms in human genomes (Choo, 2007). The high number of polymorphisms belonging to this region makes its genotyping quite challenging because of the continuous discoveries of new HLA alleles (Stefani et al., 2022; Naito and Okada, 2022).

All known HLA alleles are reported in the IPD-IMGT/HLA database (Robinson et al., 2013); so far, HLA-A and HLA-C present more than 4000 known alleles each, and HLA-B reports over 5000 different alleles (according to the IPD-IMGT/HLA database release 3.50). Therefore, many of these alleles have a very low frequency in the worldwide population (Cook et al., 2021). It is noteworthy that the complexity of the HLA region depends on the huge number of variants mapping in the regions and that the variants sites might have multiple complex alleles (indels and/or more than one single nucleotide).

Thanks to several bioinformatic strategies, it was possible to investigate short genomic sequences coming from whole genome sequencing (WGS) or whole exome sequencing (WES), pointing out the different HLA types (Erlich, 2015). The knowledge of HLA variants and alleles makes it possible to infer HLA patterns. However, the standard WGS-based imputation estimates missing genotypes of the HLA region inaccurately (Uffelmann et al., 2021).

With the advent of the Human Genome Project (Collins and Fink, 1995), the MHC region started to be unraveled, stating its complexity (Meyer and Nunes, 2017; Kennedy et al., 2017). Many GWAS for complex traits, such as cardiovascular, metabolic, and neurological diseases, have reported several associations with markers of the MHC region (Kennedy et al., 2017; Naito and Okada, 2022).

The extreme complexity of this region always makes imputation quite difficult (Uffelmann et al., 2021). Due to the high levels of linkage disequilibrium (several extended haplogroups have been identified) and the great number of polymorphisms, it is important that the imputation process can use a good reference panel containing as many haplotypes as possible (Erlich, 2015; Meyer and Nunes, 2017; Cook et al., 2021).

One of the major challenges to successfully performing genotype imputation in this complex region is to find a good and detailed reference panel. Thus, panels specific to the MHC/HLA region need to be accurately fine-mapped and comprise as many samples as possible, to overcome the enormous number of polymorphisms (Cook et al., 2021). When using sequencing, reliable imputation of the HLA region does need high read (short sequences) coverage and depth to call all the variability of the HLA region (Douillard et al., 2021).

Several tools for imputation on HLA can infer missing HLA genotypes based on reference datasets and/or individual SNPs (Douillard et al., 2021).

Over the years, many imputation algorithms able to tackle LD features of the MHC region were developed (Meyer and Nunes, 2017; Naito and Okada, 2022). All the algorithms for HLA imputation are based on probabilistic approaches. The most known and used are HLA*IMP (Dilthey et al., 2011), specific to the European population, which was subsequently improved in HLA*IMP:02 (Dilthey et al., 2013), able to handle multiple populations; indeed, SNP2HLA (Jia et al., 2013), based on Beagle, was implemented to impute not only the HLA-specific alleles but also the related aminoacidic sequence; finally, CookHLA (Cook et al., 2021) improved SNP2HLA algorithm, accounting better for linkage disequilibrium. Due to the complexity of studying HLA regions with imputed data, researchers developed HLA-TAPAS (Luo et al., 2021), a three-in-one solution for reference panel construction, imputation, and association studies on HLA genetic data.

As for all the genome regions and populations, imputation accuracy is influenced by the choice of the reference panel. At first, the HLA-specific reference panel was built for a single ancestry, considering the haplotype structures (due to the linkage disequilibrium in the region) of the European population (Dilthey et al., 2011). Moreover, the advent of NGS technologies boosted the construction of larger HLA-specific reference panels, leading to a more accurate investigation between HLA genotypes and diseases (Naito and Okada, 2022). Most recent research suggested that using multi-ethnic reference panels (Luo et al., 2021) could guarantee a more accurate genotype imputation (Meyer and Nunes, 2017; Douillard et al., 2021; Naito and Okada, 2022). A multi-ethnic reference panel was built from a high-coverage WGS dataset (Luo et al., 2021) and made available on the Michigan Imputation Server, together with an optimized procedure for HLA imputation. However, the accuracy and the sensitivity of the method are a major concern, as mixed populations can still lead to inaccurate results even if better than the ones using the standard reference panels (Douillard et al., 2021).