Modern machine learning methods are designed to exploit complex patterns in data regardless of their form, while not necessarily revealing them to the investigator. Here we demonstrate situations where modern machine learning methods are ill-equipped to reveal feature interaction effects and other nonlinear relationships. We propose the use of a conjecturing machine that generates feature relationships in the form of bounds for numerical features and boolean expressions for nominal features that are ignored by machine learning algorithms. The proposed framework is demonstrated for a classification problem with an interaction effect and a nonlinear regression problem. In both settings, true underlying relationships are revealed and generalization performance improves. The framework is then applied to patient-level data regarding COVID-19 outcomes to suggest possible risk factors.