Fun-Lda: A Latent Dirichlet Allocation Model for Predicting Tissue-Specific Functional Effects of Noncoding Variation

Abstract

We describe here a new method based on a latent Dirichlet allocation model for predicting functional effects of noncoding genetic variants in a cell type and tissue specific way (FUN-LDA) by integrating diverse epigenetic annotations for specific cell types and tissues from large scale epigenomics projects such as ENCODE and Roadmap Epigenomics. Using this unsupervised approach we predict tissue-specific functional effects for every position in the human genome. We demonstrate the usefulness of our predictions using several validation experiments. Using eQTL data from several sources, including the Genotype-Tissue Expression project, the Geuvadis project and TwinsUK cohort, we show that eQTLs in specific tissues tend to be most enriched among the predicted functional variants in relevant tissues in Roadmap. We further show how these integrated functional scores can be used to derive the most likely cell/tissue type causally implicated for a complex trait using summary statistics from genome-wide association studies, and estimate a tissue-based correlation matrix of various complex traits. We find large enrichment of heritability in functional components of relevant tissues for various complex traits, with FUN-LDA yielding the highest enrichment estimates relative to existing methods. Finally, using experimentally validated functional variants from the literature and variants possibly implicated in disease by previous studies, we rigorously compare FUN-LDA to state-of-the-art functional annotation methods such as GenoSkyline, ChromHMM, Segway, and IDEAS, and show that FUN-LDA has better prediction accuracy and higher resolution compared to these methods. In summary, we describe a new approach and perform rigorous comparisons with the most commonly used functional annotation methods, providing a valuable resource for the community interested in the functional annotation of noncoding variants. Scores for each position in the human genome and for each ENCODE/Roadmap tissue are available from http://www.columbia.edu/∼ii2135/funlda.html.

Identifying and exploiting trait-relevant tissues with multiple functional annotations in genome-wide association studies

“…For example, a two-step analysis procedure was proposed to first identify trait-relevant tissue group and then identify trait-relevant tissue within the tissue group [ 80 ]. In addition, using synthetic annotations generated from Genoskyline [ 28 ] or FUN-LDA [ 81 ] could be particularly useful for identifying fine scale trait-relevant tissues. Our method can be easily adapted to incorporate a two-step analysis procedure and/or accommodate synthetic annotations, and has the potential to yield better trait-tissue relevance resolution in the future.…”

Identifying gene targets for brain-related traits using transcriptomic and methylomic data from blood

“…11). Previous studies have indicated that functional variants (predicted by chromatin activity data) in enhancers are less likely to be shared across many tissues compared with those in promoters 26,27 , and that cell type-specific eQTLs are more dispersedly distributed around the transcription start site than eQTLs affected expression in multiple cell types 28,29 . These results seem to suggest that tissuespecific eQTLs are enriched in distal regulatory elements (i.e.…”

A semi-supervised approach for predicting cell-type specific functional consequences of non-coding variation using MPRAs

“… are the labels for m variants with MPRA validated labels; are the predicted values for a large number ( l ) of variants from a prior unsupervised method. We adopt FUN-LDA score in the current GenoNet algorithm because it is one of only a handful of tissue-specific functional scores available genome-wide, recognizing that other unsupervised scores can be readily incorporated into GenoNet in the future 18 ; X i are the functional annotations; γ I is a tuning parameter that controls how the unlabeled data are being used. When γ I = 0, the method is fully supervised; when γ I = ∞, the method is fully unsupervised (Supplementary Table 1 ).…”

“…We adopt Elastic-net because of its superior performance when the features are correlated and have sparse non-zero coefficients 21 ; are the labels for m variants with MPRA validated labels; are the predicted values for a large number ( l ) of variants from a prior unsupervised method. We choose the FUN-LDA score in the current GenoNet implementation because it is one of only a handful of tissue-specific functional scores available genome-wide, recognizing that other unsupervised scores can be readily incorporated into GenoNet in the future 18 . γ I is a tuning parameter that controls how the unlabeled data are being used 18 .…”