Data-driven approaches to genomic discovery have been accelerated by emerging efforts in machine learning. However, due to the inherent complexity of genomic data, it can be challenging to model or utilize the data and their intricate relationships. In this work, we propose a framework for genomic prediction utilizing information from various genomic databases. We use a knowledge graph following existing work to extract gene representations and either use XGBoost or construct a graph to rank feature importance. By filtering key features and computing relevancy scores with genes that are known to be associated or unassociated with a specified area, we recommend unlabeled gene candidates with a high likelihood of association for further genomic research. We demonstrate how this framework works by applying it to autophagy genomics, illustrating its potential as a powerful recommendation system for genomic discovery.