rvTWAS: identifying gene-trait association using sequences by utilizing transcriptome-directed feature selection

Abstract

Towards the identification of genetic basis of complex traits, transcriptome-wide association study (TWAS) is successful in integrating transcriptome data. However, TWAS is only applicable for common variants, excluding rare variants in exome or whole genome sequences. This is partly because of the inherent limitation of TWAS protocols that rely on predicting gene expressions. Briefly, a typical TWAS protocol has two steps: it trains an expression prediction model in a reference dataset containing gene expressions and genotype, and then applies this prediction model to a genotype-phenotype dataset to impute the unobserved expression (that is called GReX) to be associated to the phenotype. In this procedure, rare variants are not used due to its low power in predicting expressions. Our previous research has revealed the insight into TWAS: the two steps are essentially genetic feature selection and aggregations that do not have to involve predictions. Based on this insight disentangling TWAS, the inability of using rare variants to predict expression traits is no longer an obstacle. Herein, we developed rare variant TWAS, or rvTWAS, that first uses a Bayesian model to conduct expression-directed feature selection and then use a kernel machine to carry out feature aggregation, forming a model leveraging expressions for association mapping including rare variants. We demonstrated the performance of rvTWAS by thorough simulations and real data analysis in three psychiatric disorders, namely schizophrenia, bipolar disorder, and autism spectrum disorder. rvTWAS will open a door for sequence-based association mappings integrating gene expressions.

Autoencoder-transformed transcriptome improves genotype-phenotype association studies

“…First, our evaluation focused solely on the performance using the BSLMM method for GReX prediction. However, our AE-TWAS framework can also be adapted to other TWAS methods, such as ElasticNet implemented by PrediXcan and kernel methods articulated by ourselves (11,15,16,32). We chose BSLMM because our previous experiences show its supremacy to ElasticNet (16) and popularity in the field.…”