Many promising drug candidates fail because they are broken down too quickly by the body before they can have a therapeutic effect. Predicting metabolic stability early in the drug development process is therefore an important challenge, but available experimental data is often limited. In this project, I explored whether machine learning models could improve these predictions by learning meaningful representations of molecular structure from millions of compounds. Using variational autoencoders and chemical fingerprints, I developed a framework that provides additional context for data-sparse property prediction and helps identify molecules with favorable metabolic stability.