TABICO: Transformer Architecture with Bayesian Inference for Cascading Omics

초록

In the field of biology, numerous emerging methodologies and analytical approaches are uncovering implicit associations among multiple domains. These interconnected branches of biological study, collectively known as multiomics, include genomics, transcriptomics, phenomics, proteomics, and more. To achieve the goal of multi-omics integration, we propose a novel deep learning approach called Transformer Architecture with Bayesian Inference for Cascading Omics (TABICO). This method aims to provide comprehensive interpretations by integrating observations measured through various methodologies. The core concept of TABICO is that the attention mechanism can be probabilistically interpreted, and the distribution of observations with multiple (random) variables can be modeled by chaining the attention with variational inference. Thus, TABICO leverages the attention mechanism to understand the intricate relationships and dependencies among the various biological domains.