Knowledge Graphs for Explainable Artificial Intelligence and Decision-Transparency in Genomics: A Review of Methods, Applications, and Challenges

Abstract

In recent years, the field of Artificial Intelligence has emerged as one of the fastest growing and extensively used technologies in the scientific world, and is transforming the way large scale complex data is processed and understood. In genomics, a field with a ton of different and massive datasets, AI has become an increasingly important tool for speeding up the discovery process, but there are concerns with this and transparency, as well as with the trust AI can bring. To address this need, we performed a structured literature review of the few existing systems, explainability methods, and unaddressed challenges that use genomic knowledge graphs and explainable or comprehensible AI, and compiled a pipeline for the field.The amount of genetic data produced is expanding at a rate far more rapid than can be processed. While this data has great potential to enhance precision medicine, traditional analytical approaches struggle to effectively address the volume and types of sequence variants, gene annotations, multi-omics profiles, and complex regulatory networks. One of the more useful approaches to solve this problem is the use of Knowledge Graphs (KGs), which structure biological information as a network of entities and relationships. These graphs can be given a uniform language by using ontologies.The challenge is that many of the AI models that are created on these knowledge graphs remain as 'black boxes'. They make predictions without giving reasons behind their predictions, which is a major drawback in a discipline such as genomics, where decisions can have direct impact on patient care and must stand up to clinical and regulatory scrutiny. The study explores the efforts being made to bridge that gap. We discuss the most prominent KG systems currently available, explore strategies under development to increase the interpretability of AI predictions and assess the remaining challenges.