Improved genetic risk scoring algorithm for type 1 diabetes prediction

Abstract

Precise risk prediction of type 1 diabetes (T1D) facilitates early intervention and identification of risk factors prior to irreversible beta-islet cell destruction, and can significantly improve T1D prevention and clinical care. Sharp et al. (2019) developed a genetic risk scoring (GRS) system for T1D (T1D-GRS2) capable of predicting T1D risk in children of European ancestry. The T1D-GRS2 was developed on the basis of causal genetic variants, thus may be applicable to minor populations, while a trans-ethnic GRS for T1D may avoid the exacerbation of health disparities due to the lack of genomic information in minorities. Here, we describe a T1D-GRS2 calculator validated in two independent cohorts, including African American (AA) children and European American (EA) children. Participants were recruited by the Center for Applied Genomics (CAG) at the Children’s Hospital of Philadelphia (CHOP). It demonstrates that GRS2 is applicable to the T1D risk prediction in the AA cohort, while population-specific thresholds are needed for different populations. The study highlights the potential to further improve T1D-GRS2 performance with the inclusion of additional genetic markers.

Introduction

Type 1 diabetes (T1D), which is caused by autoimmune destruction of pancreatic β-cells, is most prevalent in individuals of European ancestry, but also presents a serious burden among individuals of African ancestry¹. Once diagnosed, the disease progress is irreversible, and patients will require lifelong insulin therapy. Precise T1D risk prediction is required to support preventative studies where intervention in advance of pancreatic β-cells destruction has enormous therapeutic potential. The genetic risk scoring (GRS) system for T1D, developed by Sharp et al. (2019), T1D-GRS2, uses 67 SNPs from known autoimmune loci, and is demonstrably capable of predicting T1D in children of European ancestry². To further explore the clinical potential of T1D-GRS2, we developed a computer code written in Python, which enables the calculation of T1D-GRS2 based on genotyping data generated by Illumina SNP arrays. By assessing performance in children of African as well as European ancestry, this study aims to increase the precision and broaden the scope of T1D-GRS2, both as a tool with immediate clinical application and driver of future translational studies.

Research Design and Methods

Computer code for T1D-GRS2 scoring

The T1D-GRS2 scoring was developed by Sharp et al² with 67 SNPs from known autoimmune loci, including 35 SNPs from the HLA region, and 32 SNPs from 31 non-HLA T1D susceptibility loci (Supplementary Table 1). Our code takes the input of the genetic information from a set of PLINK files containing the 67 GRS2 SNPs. The original genotyping data are from the Illumina Genotyping BeadChip with at least 550,000 SNPs genotyped. With the original genotyping data, genome-wide imputation is performed using the TOPMed Imputation Server (https://imputation.biodatacatalyst.nhlbi.nih.gov/#!) harboring the TOPMed (Version R2 on GRC38) Reference Panel. Of the 67 GRS2 SNPs, 65 can be imputed with the quality filter R² >0.3 (average R²=0.885±0.152). Two SNPs, rs144530872 corresponding to the HLA-A*2902 allele, and rs540653847 corresponding to the HLA_B_3906 allele, respectively, need to be imputed using the SNP2HLA software³. In addition, the user has the option to include four more SNPs to calculate GRS2’ (HLA-DQ: rs9273363; non-HLA: rs926169, rs10788599, and rs56380902) validated for T1D association in both EAs and AAs, with their roles in T1D risk prediction in AAs demonstrated previously ⁴. Genotyping information for these four SNPs is available in the TOPMed imputation results. The publicly available code for calculating the T1D GRS2 score was written in Bash and Python, and is available on GitHub (https://github.com/jingcqu/GRS2). The collected datasheet on the T1D GRS2 by Marc Vaudel (https://github.com/mvaudel/diabetesRiskScores/blob/master/resources/scores/T1D-GRS2) is referred to in our study with corrections.

Assessment of the performance of T1D-GRS2 in European American (EA) and African American (AA) children

Subjects: Two population samples were investigated in this study, including (1) 168 T1D AA cases versus 1366 non-diabetes AA controls; (2) 361 EA T1D cases versus 1943 non-diabetes EA controls (Table 1). Both cohorts were recruited between 2006 and 2020 by the Center for Applied Genomics (CAG) at the Children’s Hospital of Philadelphia (CHOP), which has established a large pediatric biobank coupled to comprehensive electronic medical record (EMR). Each individual was genotyped with an Illumina Genotyping BeadChip with at least 550,000 SNPs genotyped.

Table 1.

General information of the two population samples

Population		Cases	Controls	P value
AA	N	168	1366
	Male	89 (53.0%)	662 (48.5%)	0.269
	Female	79 (47.0%)	704 (51.5%)
	Age	13.6 ± 5.1	14.3 ± 3.2	0.012
	GRS2	8.13 ± 2.33	5.24 ± 2.32	1.23 × 10−48
	GRS2'	9.77 ± 3.31	5.70 ± 2.71	2.25 × 10−65
EA	N	361	1943
	Male	184 (51.0%)	990 (51.0%)	0.995
	Female	177 (49.0%)	953 (49.0%)
	Age	12.7 ± 4.5	13.5 ± 3.4	9.15 × 10−5
	GRS2	10.52 ± 2.20	7.41 ± 2.53	2.37 × 10−96
	GRS2'	12.70 ± 3.04	8.37 ± 3.10	1.17 × 10−117

Data analysis

The population ancestry of each individual was both self-reported and validated by principal component analysis (PCA) with genome-wide SNP markers. The GRS2 and the GRS2’ (with four additional SNPs associated with T1D in African Population) were calculated for each subject. The GRS2 and the GRS2’ were compared between different groups with independent t-test using IBM SPSS Statistics Version 23 software. The GRS scores were assessed for their predictive performance in each population by the Area Under the ROC Curve (AUC).

Results

Lower GRS2 and GRS2’ in the AA population

As shown in Table 1 and Table 2, a significant difference was detected between T1D cases and controls for both GRS2 and GRS2’ in both AA and EA cohorts, which suggests the feasibility of T1D prediction by GRS2 and/or GRS2’. T1D AA cases had lower GRS2 and GRS2’ scores than the EA cases, and AA controls had lower GRS2 and GRS2’ scores than the EA controls. These findings suggest population-specific thresholds of GRS2 and GRS2’ are needed for AA and EA populations.

Table 2.

P values of comparisons between different groups

	AA Cases	AA Controls	EA Cases	EA Controls
GRS
AA cases	-	1.23 × 10−48	3.62 × 10−27	-
AA controls	1.23 × 10−48	-	-	8.86 × 10−127
EA cases	3.62 × 10−27	-	-	2.37 × 10−96
EA controls	-	8.86 × 10−127	2.37 × 10−96	-
GRS2
AA cases	-	2.25 × 10−65	8.98 × 10−22	-
AA controls	2.25 × 10−65	-	-	8.69 × 10−133
EA cases	8.98 × 10−22	-	-	1.17 × 10−117
EA controls	-	8.69 × 10−133	1.17 × 10−117	-

ROC analysis of GRS2 and GRS2’ in AA and EA populations

Consequently, we performed ROC analysis in both AA and EA populations. The GRS2 had an AUC (95% CI) of 0.807 (0.779, 0.835) to predict T1D in the CAG AA cohort, compared to AUC (95% CI) of 0.823 (0.804, 0.842) in the CAG EA cohort. The prediction of T1D has a sensitivity of 0.613 and a specificity of 0.834 with the maximum Matthews correlation coefficient (MCC) at the cutoff of GRS2 at 7.43 in AA (Supplementary Table 2), compared to a sensitivity of 0.623 and a specificity of 0.833 at the cutoff of GRS2 = 9.75 in EA (Supplementary Table 3). These results suggest that the T1D GRS2 is applicable to both AA and EA populations.

The GRS2 performance improved in both AA and EA by including the four additional SNPs of T1D association in both African and European populations (HLA-DQ: rs9273363; non-HLA: rs926169, rs10788599, and rs56380902). The GRS2’ AUC (95% CI) = 0.826 (0.800, 0.852) improved in the CAG AA cohort (Supplementary Table 4), and the AUC (95% CI) = 0.839 (0.822, 0.857) also improved in the CAG EA cohort (Supplementary Table 5). With a specificity of 0.834, the GRS2’ has a sensitivity of 0.643 at the cutoff of GRS2’ = 8.19 in AA; a sensitivity of 0.690 at the cutoff of GRS2’ = 11.31 in EA.

Discussion

This study demonstrated that GRS2 is applicable to the T1D risk prediction in the AA cohort, though with a lower AUC score. The AA-specific GRS has been developed by Onengut-Gumuscu, et al. with demonstrated performance⁴. In the meantime, a trans-ethnic GRS for T1D may avoid the exacerbation of health disparities due to the lack of genomic information in minorities⁵. The T1D-GRS2 was developed on the basis of causal genetic variants, thus may be applicable to minor populations. Both the AA and EA individuals were recruited at the Children’s Hospital of Philadelphia (CHOP). As a limitation of this study, the sample size of recruited T1D individuals of other ethnicities was too limited for this study. Additionally, due to the required sample size and the long time period for patient recruitment, patients recruited in earlier time were mainly based on clinical diagnosis of T1D, without the results of the four T1D autoantibodies, i.e. islet cell antibodies (ICA, against cytoplasmic proteins in the β-cell), antibodies to glutamic acid decarboxylase (GAD-65), insulin autoantibodies (IAA), and IA-2A, antibodies to protein tyrosine phosphatase. However, the potential of mixed non-autoimmune pediatric diabetes might decrease the power of this study and cause bias towards false-negative results, thus won’t make our demonstration of the performance of T1D-GRS2 less convincing.

A lower GRS2 score is observed in the AA cohort, highlighting the importance of using population-specific reference values in the GRS2 for T1D risk prediction in AAs. The AA cohort has a lower prevalence of T1D than the EA cohort (0.57/1,000 vs 2.0/1,000)⁶. The lower GRS2 score observed in the AA cohort partially represents the lower genetic risk of T1D in the AA population.

As shown by this study, the performance of GRS2 in the AA population could be further improved by including four additional SNP markers associated with AA T1D. The marker rs926169 is from the CTLA4 region. CTLA4 encodes cytotoxic T-lymphocyte-associated protein 4, which transmits inhibitory signals to attenuate T-cell activation⁷. One SNP, rs3087243, has been included in the GRS2. However, a previous study has shown that more than one association signals have been seen in the CTLA4 region⁸. The additional marker, rs926169, has a low linkage disequilibrium (LD) r² = 0.162, D’ = 0.940 in the AA population, and r² =0.358, D’ = 0.816 in the EA population.

A previous study showed that the additional marker, rs9273363, at the human leukocyte antigen (HLA) DR-DQ region maps to a potential enhancer region of HLA-DQB1, and is associated with T1D by tagging the HLA DQB1*03:02 haplotype⁹. The GRS2 scoring system includes 35 SNPs from the HLA region. A SNP, rs9275490, is included in GRS2 to tag the DQA1*03:0X-DQB1*03:02 haplotype. However, the two SNPs have a low r² of 0.253, with D’= 0.987 in the AA cohort, and a low r² of 0.308, D’ = 1 in the EA population.

The marker rs10788599 is from the renalase, FAD dependent amine oxidase gene (RNLS) region. RNLS may contribute to T1D genetic susceptibility by its role in JAK-STAT signaling in immune-mediated diseases by activating STAT3¹⁰. The SNP rs60888743 has been included in the GRS2. As shown by this data, the two SNPs have a low LD r² of 0.007, with D’ = 0.148 in the AA cohort, and r² of 0.083, D’ = 0.387 in the EA cohort.

The SNP rs56380902 maps to the gasdermin B gene (GSDMB) locus. This locus at chr17q21.1 was not covered in the GRS2 scoring system. GSDMB encodes a gasdermin-domain containing protein that is involved in T cell-mediated cytotoxicity by inducing pyroptosis¹¹. The association of GSDMB and T1D has been validated in both European^12–14 and African populations⁴. Besides T1D, GSDMB has reported associations with other autoimmune diseases, such as rheumatoid arthritis¹⁵, and autoinflammatory disease, such as asthma¹⁶. This locus has not been covered in the GRS2 system.

At this time, the T1D-GRS2 calculation code developed in this study takes input of the genotyping imputation results. Due to the automated TOPMED imputation requiring elevated privileges, the entirety of a GRS calculation pipeline from the original genotyping BeadChip data can’t be run in one go on a high performance computing cluster. For this purpose, the T1D-GRS2 calculation based on a cloud computing platform, e.g. Amazon Web Services (AWS), to enable its clinical application, is under plan.

In conclusion, the results of this study are twofold. This study demonstrates the performance of the T1D-GRS2 calculator in both AA and EA cohorts, which implies that GRS2 may be applicable to individuals of other ethnicities or mixed ethnicity. On the other hand, the results of this study demonstrate that GRS2 improves in both AA and EA populations with the inclusion of four additional targeted SNPs, HLA-DQ: rs9273363; non-HLA: rs926169, rs10788599, and rs56380902 (noted by GRS2’). In addition, the lower GRS2 scores observed in AA individuals highlight population-specific reference values in GRS2 with consequences for T1D risk prediction in the AA population. The performance, as well as population-specific reference values, of the T1D-GRS2 in other ethnicities warrants for further study.

Data availability statement:

The data that support the findings of this study are available from the corresponding author upon reasonable request.