Limits...
Empirical evaluations of analytical issues arising from predicting HLA alleles using multiple SNPs.

Zhang XC, Li SS, Wang H, Hansen JA, Zhao LP - BMC Genet. (2011)

Bottom Line: Specifically, we have found that utilizing imputed in addition to genotyped SNPs generally yields comparable if not better performance in prediction accuracies.Further, when the training set includes multi-ethnic populations, the resulting models are reliable and perform well for the same subpopulations across all HLA genes.In contrast, the predictive models built from single ethnic populations have superior performance within the same ethnic population, but are not likely to perform well in other ethnic populations.

View Article: PubMed Central - HTML - PubMed

Affiliation: Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, 1100 Fairview Ave N, Seattle, WA 98109, USA.

ABSTRACT

Background: Numerous immune-mediated diseases have been associated with the class I and II HLA genes located within the major histocompatibility complex (MHC) consisting of highly polymorphic alleles encoded by the HLA-A, -B, -C, -DRB1, -DQB1 and -DPB1 loci. Genotyping for HLA alleles is complex and relatively expensive. Recent studies have demonstrated the feasibility of predicting HLA alleles, using MHC SNPs inside and outside of HLA that are typically included in SNP arrays and are commonly available in genome-wide association studies (GWAS). We have recently described a novel method that is complementary to the previous methods, for accurately predicting HLA alleles using unphased flanking SNPs genotypes. In this manuscript, we address several practical issues relevant to the application of this methodology.

Results: Applying this new methodology to three large independent study cohorts, we have evaluated the performance of the predictive models in ethnically diverse populations. Specifically, we have found that utilizing imputed in addition to genotyped SNPs generally yields comparable if not better performance in prediction accuracies. Our evaluation also supports the idea that predictive models trained on one population are transferable to other populations of the same ethnicity. Further, when the training set includes multi-ethnic populations, the resulting models are reliable and perform well for the same subpopulations across all HLA genes. In contrast, the predictive models built from single ethnic populations have superior performance within the same ethnic population, but are not likely to perform well in other ethnic populations.

Conclusions: The empirical explorations reported here provide further evidence in support of the application of this approach for predicting HLA alleles with GWAS-derived SNP data. Utilizing all available samples, we have built "state of the art" predictive models for HLA-A, -B, -C, -DRB1, -DQB1 and -DPB1. The HLA allele predictive models, along with the program used to carry out the prediction, are available on our website.

Show MeSH
Comparison of prediction accuracies between models built with and without imputed SNPs from four arrays. Half of the common set of samples genotyped on Affy 500K, Affy 6.0, Illumina 550K, Illumina 1.2M arrays in the WTCCC cohort were used as the training set (N = 501) and the other half were used as the validation set (N = 500). Each panel shows a comparison of prediction accuracies for the validation set, with models built using only SNPs observed from the array or using HapMap SNPs observed and imputed from the array. The confidence threshold was set at CT = 0.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3111398&req=5

Figure 1: Comparison of prediction accuracies between models built with and without imputed SNPs from four arrays. Half of the common set of samples genotyped on Affy 500K, Affy 6.0, Illumina 550K, Illumina 1.2M arrays in the WTCCC cohort were used as the training set (N = 501) and the other half were used as the validation set (N = 500). Each panel shows a comparison of prediction accuracies for the validation set, with models built using only SNPs observed from the array or using HapMap SNPs observed and imputed from the array. The confidence threshold was set at CT = 0.

Mentions: To extend this observation beyond Affy 5.0, we used the WTCCC data, to evaluate the usefulness of imputed SNPs from Affy 500K, Affy 6.0, Illumina 550K, and Illumina 1.2M. Figure 1 has four panels of figures, labeled by the corresponding technologies. Comparing accuracies between models built with and without imputed SNPs, one would conclude that predictive models built with imputed SNPs again have either comparable or better accuracies, with few exceptions. For those exceptions, losses of accuracies are less than 1%. Certainly, variations between two classes of models are much less than those between intermediate and high resolutions. Details in accuracies of prediction across different CTs are listed in the Additional file 1: Table S3.


Empirical evaluations of analytical issues arising from predicting HLA alleles using multiple SNPs.

Zhang XC, Li SS, Wang H, Hansen JA, Zhao LP - BMC Genet. (2011)

Comparison of prediction accuracies between models built with and without imputed SNPs from four arrays. Half of the common set of samples genotyped on Affy 500K, Affy 6.0, Illumina 550K, Illumina 1.2M arrays in the WTCCC cohort were used as the training set (N = 501) and the other half were used as the validation set (N = 500). Each panel shows a comparison of prediction accuracies for the validation set, with models built using only SNPs observed from the array or using HapMap SNPs observed and imputed from the array. The confidence threshold was set at CT = 0.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3111398&req=5

Figure 1: Comparison of prediction accuracies between models built with and without imputed SNPs from four arrays. Half of the common set of samples genotyped on Affy 500K, Affy 6.0, Illumina 550K, Illumina 1.2M arrays in the WTCCC cohort were used as the training set (N = 501) and the other half were used as the validation set (N = 500). Each panel shows a comparison of prediction accuracies for the validation set, with models built using only SNPs observed from the array or using HapMap SNPs observed and imputed from the array. The confidence threshold was set at CT = 0.
Mentions: To extend this observation beyond Affy 5.0, we used the WTCCC data, to evaluate the usefulness of imputed SNPs from Affy 500K, Affy 6.0, Illumina 550K, and Illumina 1.2M. Figure 1 has four panels of figures, labeled by the corresponding technologies. Comparing accuracies between models built with and without imputed SNPs, one would conclude that predictive models built with imputed SNPs again have either comparable or better accuracies, with few exceptions. For those exceptions, losses of accuracies are less than 1%. Certainly, variations between two classes of models are much less than those between intermediate and high resolutions. Details in accuracies of prediction across different CTs are listed in the Additional file 1: Table S3.

Bottom Line: Specifically, we have found that utilizing imputed in addition to genotyped SNPs generally yields comparable if not better performance in prediction accuracies.Further, when the training set includes multi-ethnic populations, the resulting models are reliable and perform well for the same subpopulations across all HLA genes.In contrast, the predictive models built from single ethnic populations have superior performance within the same ethnic population, but are not likely to perform well in other ethnic populations.

View Article: PubMed Central - HTML - PubMed

Affiliation: Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, 1100 Fairview Ave N, Seattle, WA 98109, USA.

ABSTRACT

Background: Numerous immune-mediated diseases have been associated with the class I and II HLA genes located within the major histocompatibility complex (MHC) consisting of highly polymorphic alleles encoded by the HLA-A, -B, -C, -DRB1, -DQB1 and -DPB1 loci. Genotyping for HLA alleles is complex and relatively expensive. Recent studies have demonstrated the feasibility of predicting HLA alleles, using MHC SNPs inside and outside of HLA that are typically included in SNP arrays and are commonly available in genome-wide association studies (GWAS). We have recently described a novel method that is complementary to the previous methods, for accurately predicting HLA alleles using unphased flanking SNPs genotypes. In this manuscript, we address several practical issues relevant to the application of this methodology.

Results: Applying this new methodology to three large independent study cohorts, we have evaluated the performance of the predictive models in ethnically diverse populations. Specifically, we have found that utilizing imputed in addition to genotyped SNPs generally yields comparable if not better performance in prediction accuracies. Our evaluation also supports the idea that predictive models trained on one population are transferable to other populations of the same ethnicity. Further, when the training set includes multi-ethnic populations, the resulting models are reliable and perform well for the same subpopulations across all HLA genes. In contrast, the predictive models built from single ethnic populations have superior performance within the same ethnic population, but are not likely to perform well in other ethnic populations.

Conclusions: The empirical explorations reported here provide further evidence in support of the application of this approach for predicting HLA alleles with GWAS-derived SNP data. Utilizing all available samples, we have built "state of the art" predictive models for HLA-A, -B, -C, -DRB1, -DQB1 and -DPB1. The HLA allele predictive models, along with the program used to carry out the prediction, are available on our website.

Show MeSH