Organ-specific prioritization and annotation of non-coding regulatory variants in the human genome

Abstract

Identifying non-coding regulatory variants in the human genome remains a challenging task in genomics. Recently we advanced our leading regulatory variant database, RegulomeDB, to its second version. Building upon this comprehensive database, we developed a novel machine-learning architecture with stacked generalization, TLand, which utilizes RegulomeDB-derived features to predict regulatory variants at cell or organ-specific levels. In our holdout benchmarking, TLand consistently outperformed state-of-the-art models, demonstrating its ability to generalize to new cell lines or organs. We trained three types of organ-specific TLand models to overcome the common model bias toward high data availability cell lines or organs. These models accurately prioritize relevant organs for 2 million GWAS SNPs associated with GWAS traits. Moreover, our analysis of top-scoring variants in specific organ models showed a high enrichment of relevant GWAS traits. We expect that TLand and RegulomeDB will further advance our ability to understand human regulatory variants genome-wide.

Enhancing Portability of Trans-Ancestral Polygenic Risk Scores through Tissue-Specific Functional Genomic Data Integration

“…The goal of this model is to identify genomic regions with more significant regulatory effect within a specific tissue over organism-wide predictions. Finally, we also included the recently developed TLand model, which uses a stacked generalization model to learn RegulomeDB-derived features across all ENCODE hg38 experiments and predictions from Sei 15 to assign probabilities for regulatory variants on a cell-type and organ-specific level for 51 tissue-specific models 18 .…”

“…The fourth is TLand, which constructs a stacked generalization model to learn RegulomeDB-derived features to predict regulatory variants at a cell-specific level or organ-specific level. Further details on the TLand model are reported in Zhao et al 2023.…”

“…Here we introduce the TITR algorithm, a trans-ancestral PRS model that leverages SNP-level regulatory probabilistic scores from RegulomeDB. Three models - SURF 16 , TURF 17 and TLand 18 -capture experimental evidence from the ENCODE project on either an organism-wide or tissue-specific scale and also feature allele-specific predictions from DeepSEA and Sei. We show that, by highlighting mutations with the greatest degree of organism-wide or tissue-specific regulatory evidence, we capture a significant proportion of SNP heritability and achieve greater predictive accuracy for trans-ancestral PRS compared to other published methods.…”