Identifying novel genes associated with autophagy (ATG) in man remains an important task for gaining complete understanding on this fundamental physiological process. A machine-learning guided approach can highlight potentially "missing pieces" linking core autophagy genes with understudied, "dark" genes that can help us gain deeper insight into these processes. In this study, we used a set of 103 (out of 288 genes from the Autophagy Database, ATGdb), based on the presence of ATG-associated terms annotated from 3 secondary sources: GO (gene ontology), KEGG pathway and UniProt keywords, respectively. We regarded these as additional confirmation for their importance in ATG. As negative labels, we used the OMIM list of genes associated with monogenic diseases (after excluding the 288 ATG-associated genes). Data associated with these genes from 17 different public sources were compiled and used to derive a Meta Path/XGBoost (MPxgb) machine learning model trained to distinguish ATG and non-ATG genes (10-fold cross-validated, 100-times randomized models, median AUC = 0.994 +/-0.0084). Sixteen ATG-relevant variables explain 64% of the total model gain, and 23% of the top 251 predicted genes are annotated in ATGdb. Another 15 genes have potential ATG associations, whereas 193 do not. We suggest that some of these 193 genes may represent "autophagy dark genes", and argue that machine learning can be used to guide autophagy research in order to gain a more complete functional and pathway annotation of this complex process.