309-OR: Deep Learning Model of Pancreatic Islet Epigenome Refines Association Signals at Type 2 Diabetes Susceptibility Loci

Abstract

Translation of genome-wide association study findings is hampered by challenges in pinpointing the causal variant and interpretation of the biological significance of noncoding variation. Recently, deep learning models have been successfully applied to study variant effects on DNA accessibility. Here we collected 30 islet-specific publicly available epigenomic datasets, including measures of DNA accessibility, methylation, histone marks, and transcription factor (TF) binding, and used them to train a convolutional neural network (CNN). The resulting model of pancreatic islet epigenome achieved mean AUC of 0.847 per predicted feature (range: 0.713-0.976), with best predictive power for features marking promoters, DNA accessibility and TF binding. Convolution filters in the first CNN layer discovered binding motifs of several TFs with known islet functions, including FOXA2, HNF1A and RFX6. We used this model to predict islet regulatory effects of variants from a recent T2D GWAS study reporting 403 T2D-risk signals in ~900,000 individuals of European ancestry. We observed overall convergence of three complementary approaches to prioritize functional variants: genetic fine-mapping, regulatory annotation enrichment and islet CNN model predictions. We demonstrate the value of integrating deep learning approaches with fine-mapping to identify causal variants by highlighting an association signal at the PROX1 locus, fine-mapped to two nearby plausible variants. Both variants sit in an islet strong open enhancer, and both exhibit strong open chromatin allelic imbalance, but the CNN model predicts only rs17712208 to have regulatory impact in islets (p=1.69e-160), with the A allele disrupting an HNF1B binding motif, resulting in loss of the H3K27ac mark, indicative of an active regulatory element. These predictions will facilitate further functional follow-up studies to fully elucidate the underlying disease mechanisms at prioritized variants. Disclosure A. Wesolowska-Andersen: None. M. Thurner: None. A. Mahajan: None. F. Abaitua: None. J. Torres: None. V. Nylander: None. A.L. Gloyn: Consultant; Spouse/Partner; Eli Lilly and Company, Merck & Co., Inc. Consultant; Self; Merck & Co., Inc. Consultant; Spouse/Partner; Novo Nordisk A/S, Pfizer Inc. Research Support; Self; Novo Nordisk A/S. Speaker's Bureau; Self; Novo Nordisk A/S. Other Relationship; Self; Diabetes UK, European Foundation for the Study of Diabetes. M. McCarthy: Advisory Panel; Self; European Association for the Study of Diabetes, Pfizer Inc. Consultant; Self; Eli Lilly and Company, Merck & Co., Inc. Consultant; Spouse/Partner; Merck & Co., Inc. Research Support; Self; AbbVie Inc., Boehringer Ingelheim International GmbH. Research Support; Spouse/Partner; Diabetes UK. Research Support; Self; Janssen Pharmaceuticals, Inc., Merck & Co., Inc., National Institutes of Health. Research Support; Spouse/Partner; National Institutes of Health. Research Support; Self; Novo Nordisk A/S. Research Support; Spouse/Partner; Novo Nordisk A/S. Research Support; Self; Novo Nordisk Foundation, Roche Pharma, Sanofi-Aventis, Servier, Takeda Pharmaceutical Company Limited.