188-LB: Deep Learning–Based Population Screening of Type 1 Diabetes and Celiac Disease Genetic Risk from Blood Spots at Birth

Abstract

Introduction: Screening for type 1 diabetes (T1D) genetic risk in early life can prevent life threatening complications such as diabetic ketoacidosis and allow cost-effective recruitment into intervention and immunotherapy trials. Celiac disease (CD) frequently presents as a comorbidity in those with T1D and shares a strong genetic basis with T1D. Single nucleotide polymorphism (SNP) based genetic risk scores (GRS) for both T1D and CD have shown to be very predictive of future disease but are difficult to generate and typically require SNP array genotyping. We aimed to develop a combined GRS screening panel that could be genotyped from a single dried blood spot at birth. Methods: We developed assays for proxy SNPs of common HLA-DQ haplotypes reported in previous GRS and additional loci from the most recent available genome-wide association studies (GWAS) . We used backwards stepwise regression to identify a subset of variants able to be genotyped with DNA eluted from 800 6mm dried blood spots. We developed and validated neural network models to quantify genetic risk of both T1D and CD using T1DGC and UK Celiac case-control SNP array data, validated in UK Biobank. Assays were developed with LGC Genomics and validated on 675 Seattle area samples. Results: The complete panel consisted of 71 validated SNP assays, including 11 backup variants for key loci. We generated neural nework models which demonstrate equivalent or greater predictive power (AUC: T1D=0.914, CD=0.893) to previously published GRS yet require much less expertise to apply. We have developed an algorithm for preparation of raw genotyping data and subsequent generation of GRS requiring little expertise to apply. Conclusion: A 71 SNP blood-spot screening panel is highly effective at screening genetic risk associated with T1D and CD at birth. Using a neural network model the panel enables widely available, easy to generate and inexpensive population screening of genetic risk for T1D and CD. Disclosure S. A. Sharp: None. J. M. Locke: None. Y. Xu: None. D. P. Fraser: None. L. A. Ferrat: None. M. N. Weedon: None. M. Inouye: None. R. A. Oram: Consultant; Janssen Research & Development, LLC, Research Support; Randox R & D. W. Hagopian: Research Support; Janssen Research & Development, LLC. Funding Diabetes UK (16/0005529) JDRF (3-SRA-2019-827-S-B)