Nearly all computational methods in the CADD field depend on parameters whose values are derived from various types of experimental data. Here, we draw a distinction between two broad classes of approaches: 1) where each model parameter is tied to a relevant and physically meaningful phenomenon (“Physical ML”); and 2) where model parameters may have no obvious linkage to specific physical quantities (“Pure ML”). Contrasts will be drawn between the Physical and Pure ML approaches using specific methods and applications for affinity prediction, energy estimation, and molecular docking. Both types of approaches have compelling and effective use cases. However, the requirements for data and the challenges in building and applying prediction models are quite different, and this has a dramatic effect on the best areas for applying Pure vs. Physical machine-learning. ML modelling approaches containing dozens to millions of parameters will be characterised in terms of their appropriateness, applicability, and performance for predictive modelling in drug discovery and optimisation.