Transcriptome-wide association study (TWAS) has been successful in identifying genetic basis of complex diseases. In a typical TWAS framework, one first trains a gene expression predictive model using genotype; and then in a genotype-disease cohort, one predicts the gene expression and associates the predicted with phenotype. However, rare variants (defined as their population minor allele frequency < 0.5%), which contain >90% sites of genetic diversity and proven to be functional critical, cannot be included, due to the rare variants’ inability to predict. Herein, based upon our previous work disentangling feature selection and feature aggregation in TWAS, we propose rvTWAS, a protocol that uses a Bayesian feature selection to form a set of rare variants, which are in turn aggregated using a kernel machine to be associated to diseases. We showed the superior performance of rvTWAS in both simulated and real data and discovered additional genes underlying multiple complex diseases.